Skip to content
Projects

Lethe044/hermes-life-os has 242 GitHub stars and 16 forks. GitHot measured +1 stars in 24 hours and +6 in 7 days. Last checked 2026-11-15 13:59 UTC. Last push 2026-11-15.

How these counts are measured

What you can do

  • Memory - Stores every mood, meal, sleep entry, workout, stress log - recalls before every response
  • Skills - Life OS playbook defines daily rhythm, pattern detection rules, and briefing format
  • Cron - Automated briefings at 07:00, 12:00, 18:00, 23:00, and weekly Monday reviews
  • Gateway - Delivers briefings via terminal - extensible to Telegram, email, SMS

Measured star growth

in 24h
+1
in 7 days
+6
in 15 days
Update delayed
in 1 month
Update delayed

Net star changes over overlapping periods, not daily totals. Missing measurements are left blank. Stars measure attention, not software quality.

About

Local-first personal OS: mood, sleep, habits, tasks, day planning and a memory that connects the dots. Works with Ollama, OpenAI, Anthropic and OpenRouter.

Productivity

README.md

Original

Hermes Life OS: the personal OS that remembers, plans and grows with you

Hermes Life OS 🧠

Tests PyPI Python Versions License: MIT GitHub stars Last commit

The personal OS that grows with you. Track how you feel, plan your day, keep your life organised, and let an agent connect the dots. Local first, encrypted if you want, no account needed.

Quick start · Tool reference · Changelog · Contributing · Security

Built for the NousResearch "Show us what Hermes Agent can do" hackathon.

Quick Tour

Ask in plain language; Hermes calls the right tool and answers from your own data.

Hermes planning a day around free time and showing a life wheel

You want to Say something like Tool
Pick today's focus "my top three today are the report, the gym and calling mum" add_priority
Get a plan "plan my day" plan_my_day
Check in on your life "rate my health 7 and money 3" log_life_wheel
Keep a backlog "add Dune to my books" add_library_item
Use your own calendar app "export my calendar" export_calendar
Empty your head "remind me to look into cello lessons" add_to_inbox
Close the week "start my weekly review" start_weekly_review
Learn something for good "quiz me on my Spanish cards" review_flashcards
See the money picture "what is my net worth?" get_net_worth
Move a big goal forward "start a project: launch the website" add_project
Stop forgetting follow-ups "I am waiting for a quote from the plumber" add_waiting
Undo a mistake "put the rent task back" restore_deleted
Understand yourself "what affects my sleep?" get_correlations

All 299 tools are listed in the tool reference.

Contents

What It Does

Tell it how you feel. Log what you ate. Track your sleep. Over time it starts connecting dots you haven't: energy crashes after poor sleep, mood dips on low-hydration days, focus drops when stress spikes. Every morning it briefs you. Every evening it reflects with you. Every week it tells you what the data says about your life.

The longer you use it, the more it knows. The more it knows, the more useful it becomes.

Architecture

flowchart TD
    A([👤 You share something]) --> B
    B[🧠 REMEMBER<br/>Mood · Sleep · Meals<br/>Stress · Focus · Habits] --> C
    C[🔍 RECALL<br/>Search memory<br/>for context] --> D
    D[📊 DETECT PATTERNS<br/>Correlations across<br/>all life dimensions] --> E
    E[📋 BRIEF<br/>Personalized insight<br/>based on YOUR data] --> F
    F([🌱 Hermes knows you<br/>a little better today])

    G([⏰ Cron Schedule<br/>07:00 Morning<br/>12:00 Midday<br/>18:00 Evening<br/>23:00 Consolidate<br/>Mon 08:00 Weekly]) --> C

    style A fill:#2980b9,color:#fff
    style F fill:#27ae60,color:#fff
    style G fill:#8e44ad,color:#fff
    style D fill:#e67e22,color:#fff

Hermes Features Used

Feature How It's Used
Memory Stores every mood, meal, sleep entry, workout, stress log - recalls before every response
Skills Life OS playbook defines daily rhythm, pattern detection rules, and briefing format
Cron Automated briefings at 07:00, 12:00, 18:00, 23:00, and weekly Monday reviews
Gateway Delivers briefings via terminal - extensible to Telegram, email, SMS
Subagents Pattern detection runs across all health dimensions in parallel
Atropos RL Reward function trains Hermes to be more personal and memory-driven over time

Tracking Capabilities

Category What Hermes Tracks
🥗 Nutrition Meals, calories, protein/carbs/fat, daily totals
😴 Sleep Duration, quality score, 7-day averages
💧 Hydration Daily water intake with progress bar
💪 Fitness Workouts, duration, intensity, weekly count
🧘 Mental Stress levels, meditation sessions, gratitude logs
🎯 Focus Deep work sessions, distractions, quality scores
✅ Habits Streaks, best streaks, completion tracking
🎯 Goals Progress percentages, milestones, notes
😊 Mood & Energy Daily scores, trend detection, dip alerts
💰 Spending Expenses by category, daily/period totals
🤝 Social Time connecting with others, quality, trend
☕ Substances Caffeine, alcohol, or anything else - amount, frequency
📚 Reading Sessions, minutes, pages, titles
💊 Medication Dose taken/skipped, adherence % by medication

Pattern Detection

Hermes automatically detects and surfaces:

  • Mood dips lasting 3+ consecutive days
  • Sleep deprivation affecting focus and mood
  • Energy crashes correlated with nutrition gaps
  • Stress spikes and their triggers
  • Habit streaks worth celebrating
  • Goal stalls that need a nudge
  • Hydration gaps on high-stress days

Correlation Engine

demo/analytics.py computes real Pearson correlation coefficients between tracked metrics (mood, sleep, stress, energy, hydration) using daily-averaged values from memory. A pair is only surfaced when there's enough data (4+ overlapping days by default) and the relationship is meaningful (|r| >= 0.4). Each result reports the direction (positive/negative), strength (weak/moderate/strong), and the number of days behind it - no external dependencies required (pure Python stdlib).

Lagged (predictive) correlations: same-day correlation can't tell you whether poor sleep caused today's low mood, or whether being stressed already caused last night's poor sleep - it just says the two move together. compute_lagged_correlations() shifts one metric forward by 1-2 calendar days (correctly handling gaps from unlogged days) before correlating, which at least points the arrow of time forward: "a higher X on one day tends to be followed by a higher/lower Y N days later." Both same-day and lagged results feed every surface that already shows insights (chat replies, detect_patterns, the static and live dashboards, the weekly email) automatically, plus a dedicated get_correlation_insights tool for a deeper, on-demand analysis over a custom day range - just ask "what patterns have you noticed in my data?" or "what predicts my mood?".

Reward Function

pie title Life OS Reward Components
    "Briefing Sent - Delivered via send_briefing?" : 30
    "Memory Used - Recalled AND remembered?" : 25
    "Pattern Detected - Called detect_patterns?" : 20
    "Personalization - Referenced real context?" : 15
    "Tool Coverage - Used expected tools?" : 10

Quick Start

Works with four LLM backends - pick whichever you already have. The provider is auto-detected from whatever key is set (or force one with --provider).

pip install "hermes-life-os[all]"

# Option A - free, fully local, no API key:
ollama serve
ollama pull llama3.1

# Option B / C / D - pick one:
set ANTHROPIC_API_KEY=sk-ant-...
set OPENAI_API_KEY=sk-...
set OPENROUTER_API_KEY=sk-or-...

hermes-life-os --mode onboard
hermes-life-os --mode morning
hermes-life-os --mode chat

# force a specific backend regardless of which keys are set:
hermes-life-os --mode morning --provider anthropic

Prefer running from source instead of installing? Clone the repo and use python demo/demo_life_os.py ... in place of hermes-life-os ... above (same flags, same behavior) - see Project Structure.

Or run it with Docker - zero Python setup

# pull the pre-built image - no clone needed:
docker run --rm -it -e ANTHROPIC_API_KEY=sk-ant-... \
    -v hermes-life-os-data:/root/.hermes \
    ghcr.io/lethe044/hermes-life-os:latest --mode morning

Or build it yourself, and get a fully free trial paired with a local Ollama container (no API key at all):

git clone https://github.com/Lethe044/hermes-life-os.git
cd hermes-life-os

docker compose up -d ollama
docker compose exec ollama ollama pull llama3.1
docker compose run --rm hermes-life-os --mode onboard

All Demo Modes

Mode What Happens
onboard First-time setup - Hermes learns who you are
morning Daily briefing based on all your patterns
checkin Midday log - mood, habits, quick nudge
evening Evening reflection - wins, struggles, patterns
weekly Sunday review - what this week says about you
nutrition Log meals and get nutrition insights
sleep Log sleep and get sleep analysis
fitness Log workouts and track fitness patterns
mental Log stress, meditation, and gratitude
focus Log deep work sessions and productivity
health Full health dashboard - all data in one view
dream Dream journal - log dreams, detect patterns, sleep/stress correlation
chat Interactive conversation - type anything

Chat Mode

python demo/demo_life_os.py --mode chat

Type naturally. Hermes responds using everything it knows about you. Type exit to leave.

Example conversations:

  • "I feel stressed today, any advice?"
  • "Log my lunch - grilled chicken and rice, about 600 calories"
  • "How has my sleep been this week?"
  • "I just ran 5km, log it"
  • "What patterns are you seeing in my data?"
  • "I logged 4 hours of sleep by mistake, it was actually 7" - Hermes recalls the entry and corrects it
  • "Delete that last mood entry, I misclicked" - Hermes finds and removes it (asks for confirmation first)
  • "Set a goal to sleep 7+ hours a night" - Hermes tracks this automatically from your actual logged sleep, no manual progress updates needed
  • "How am I doing on my goals?" / "How does this week compare to last week?"
  • "Has anything been unusual lately?" / "How was I in March?"

Multi-Profile (shared households)

python demo/demo_life_os.py --mode morning --profile alex
python demo/dashboard.py --profile alex

By default everything lives at ~/.hermes/life-os/ (unchanged, single person). Passing --profile <name> (or setting LIFE_OS_PROFILE) fully isolates that person's data under ~/.hermes/life-os/profiles/<name>/ - so a household can share one install without mixing anyone's mood/sleep/ habit data. Omitting --profile always keeps working exactly as before.

Multi-User (real accounts for the local API & Slack bot)

python demo/users.py add alex --profile alex --role owner
python demo/users.py add sam  --profile sam
python demo/users.py list

Profiles isolate data; this registry is what turns a profile into a person who can log in. Each user gets their own API key (shown once, stored only as a salted PBKDF2 hash) that resolves to their own profile automatically - so a whole household or small team can share one running hermes-life-os-api server or one Slack bot, and everyone only ever sees their own data:

curl -H "X-API-Key: <alex's key>" http://127.0.0.1:8765/api/health
# {"status": "ok", "profile": "alex", "user": "alex"}

Fully opt-in and non-breaking - a single LIFE_OS_API_KEY/--profile keeps working exactly as before if you never touch users.py. Full setup, including linking a user to their Slack account, in docs/MULTI_USER.md.

Plugin System (add your own tools, no fork required)

mkdir -p ~/.hermes/life-os/plugins
cp demo/plugins_examples/dice.py ~/.hermes/life-os/plugins/
python demo/demo_life_os.py --mode chat
# "roll a d20 for me"

Drop a .py file defining a TOOLS list and a dispatch(name, inp) function into ~/.hermes/life-os/plugins/, and Hermes' LLM agent can call it like any built-in tool - next start, no core code changes. A broken plugin is skipped and reported, never crashes the app; a plugin can't shadow a built-in tool name. python demo/plugins.py lists everything currently loaded. Two ready-to-copy examples ship in demo/plugins_examples/ (a dependency-free dice/coin-flip tool, and a profile-aware screen-time tracker showing how to persist your own data). Full plugin API and a "share it with others" guide in docs/PLUGINS.md.

Encryption at Rest (optional)

set LIFE_OS_ENCRYPTION_KEY=your-passphrase-here
python demo/demo_life_os.py --mode morning

Off by default - nothing changes unless you set this. When set, every data file (profile, habits, goals, nutrition, sleep, etc.) and every line of memory.jsonl is encrypted at rest with a key derived from your passphrase (PBKDF2-HMAC-SHA256 + Fernet/AES). Existing plaintext data is read transparently and gets encrypted the next time it's written - no separate migration step. There is no password recovery - if you lose the passphrase, that data is unrecoverable by design. Requires pip install "hermes-life-os[encryption]" (or pip install cryptography if running from source).

Changing your passphrase: use hermes-life-os-rekey rather than just setting a new LIFE_OS_ENCRYPTION_KEY - the latter would leave your existing files encrypted under the old key, unreadable. The re-key tool decrypts everything with the old key, rotates the salt, and re-encrypts everything with the new one in one step (it also works to enable encryption for the first time, or disable it entirely):

hermes-life-os-backup                                    # back up first
hermes-life-os-rekey --old-key "old pass" --new-key "new pass"
hermes-life-os-rekey --new-key "new pass"                # enable for the first time
hermes-life-os-rekey --disable                           # decrypt everything back to plaintext

As of v1.38.0 re-keying covers every data store (spending, social, substance, reading, medication, achievements, templates, budgets and reminders as well as the original nine). Before that fix it only re-encrypted nine files, so after a passphrase change the others could no longer be decrypted and silently read back as empty. Backups written while a passphrase is set are encrypted too (see "Automatic Backups").

Time Blocks, Objectives, Recipes & Trips

Five more stores plus three read-only helpers, all local, per profile, encrypted with the rest of your data when encryption is on, and covered by backup, restore, export and the doctor.

"block 9 to 10:30 tomorrow for deep work"   "when am I free for 45 minutes?"
"schedule the report task this week"        "how did I spend my week?"
"objective for Q4: run 100 km"              "I ran 12 km today"
"save my lentil soup recipe"                "plan soup for dinner on Friday"
"make a shopping list from the meal plan"   "trip to Rome Nov 1 to Nov 4"
"what triggers my headaches?"               "find everything about Rome"
  • Time blocks (add_time_block, get_day_plan, update_time_block, delete_time_block, finish_time_block, find_free_time, schedule_task, get_time_allocation): put work on the clock. Overlaps are refused, the day plan shows blocks and free gaps (07:00 to 22:00), find_free_time lists the first free slots, schedule_task places a task in the first slot that fits and the allocation report shows where the hours went by category. Finishing a block can complete its task. A block starting within 15 minutes is nudged.
  • Objectives and key results (add_objective, add_key_result, record_key_result, get_okrs, update_okr, delete_okr): an objective for a week, month, quarter (2026-Q4) or year with measurable key results (upward or downward targets). Progress is compared with the share of the period that has passed: ahead of pace, on track, slightly behind or behind. Objectives ending within a week below 70% are nudged and shown in the agenda.
  • Recipes and meal plan (add_recipe, get_recipes, get_recipe, update_recipe, delete_recipe, suggest_recipes, plan_meal, get_meal_plan, clear_meals, log_planned_meal, shopping_from_plan): recipes scale to any serving count, suggest_recipes surfaces what you have not cooked lately, planned meals can repeat on several days, logging a planned meal creates the meal entry, and shopping_from_plan adds the combined ingredients to your shopping list (skipping what is already there).
  • Trips (add_trip, get_trips, get_trip, update_trip, delete_trip, add_packing_items, pack_items, remove_packing_items, add_itinerary_item, remove_itinerary_item): dates, budget, a packing list from a template (basic, documents, flight, city, beach, hiking, winter, business) and a day-by-day itinerary. A trip within three days with unpacked items is nudged; an ongoing trip shows today's plan in the agenda.
  • Symptom triggers (get_symptom_triggers): compares your symptom days with sleep, stress, water, caffeine, alcohol, nicotine and workouts of the day before, plus triggers you named yourself. Patterns in your own data, not causes and not medical advice.
  • Search everything (search_everything): one search over tasks, notes, lists, people, subscriptions, savings, debts, countdowns, supplies, symptoms, recipes, the meal plan, trips, objectives, time blocks, reminders, habits and goals, with the id to use next to every hit.
  • Closing summary (get_closing_summary): an end-of-day or end-of-week review of what got done, what was logged, what is still open and what is next.
  • Repair suggestions (get_repair_suggestions): for each problem the doctor finds in planner data, the exact tool call that fixes it. Read-only.

The proactive nudges and get_today_agenda include the new sources, and the doctor checks the five new stores (blocks, objectives, recipes and planned meals it cannot read, trips with unusable dates, shared ids).

Chores, Expiry Dates, Routines, Gift Ideas & Decisions

Five more stores and a combined calendar, all local, per profile, encrypted with the rest of your data when encryption is on, and covered by backup, restore, export, rekey and the doctor.

"water the plants every week"            "I watered them"
"my passport expires 2027-03-01"         "I renewed my insurance for 12 months"
"morning routine: water, stretch, journal"  "tick stretch"
"a poetry book for Ayse"                 "I bought the scarf"
"I chose the standing desk, review in a month"  "how did the desk decision turn out?"
"what is coming up this week?"
  • Chores (add_chore, complete_chore, get_chores, update_chore, delete_chore): anything that repeats on a schedule (every 10 days, weekly, every 3 months, yearly). The next due date counts from the day it was done, so a late chore never piles up. Overdue and due-today chores are nudged and shown in the agenda.
  • Expiry dates (add_expiry, get_expiries, renew_expiry, update_expiry, delete_expiry): a passport, a licence, insurance, a warranty. Each has a reminder window (30 days by default); inside it the item is nudged and listed in the agenda. Renewing moves the date forward (a new date, or a number of months) and keeps a history.
  • Routines (add_routine, get_routines, check_routine_step, update_routine, delete_routine): named checklists with their own schedule (daily, weekdays, weekends or chosen days). Tick steps per day (or all), see the streak of fully done days (days a routine is not scheduled do not break it) and find unfinished routines in the agenda.
  • Gift ideas (add_gift_idea, get_gift_ideas, update_gift_idea, delete_gift_idea): ideas per person with price and occasion, marked idea, bought or given. Names from your people list link to their birthday: the list shows birthdays with no gift bought yet, and a birthday within a week with saved ideas but nothing bought is nudged.
  • Decisions (add_decision, get_decisions, review_decision, update_decision, delete_decision): the options, the choice, why, what you expect and how confident you are, with a review date (30 days by default). When it arrives Hermes asks how it went; reviewed decisions show how your confidence compares with the outcomes.
  • Week ahead (get_week_ahead): one calendar for 1 to 14 days over due tasks, time blocks, planned meals, birthdays, renewals, debts, countdowns, chores, expiry dates, trips and decisions to review, with overdue items first.

search_everything now covers the five new stores, the closing summary shows chores and decisions that are waiting, and get_repair_suggestions and the doctor check them (a chore interval it cannot read, an expiry date or review date that is not a date, a routine without steps, an unknown gift status, shared ids).

Recently Deleted, Recommendation Ledger & Pinned Backups

Three safety nets, all local, per profile, encrypted with the rest of your data when encryption is on, and covered by backup, restore, export and the doctor.

"I deleted the wrong task"               "what did I delete this week?"
"put the rent task back"                 "empty the trash"
"which recommendations did I follow?"    "I followed the sleep suggestion"
"keep today's backup for good"
  • Recently deleted (get_trash, restore_deleted, empty_trash): what the planner's delete and clear tools remove is kept for 30 days. This covers whole tasks, notes, people, subscriptions, savings goals, debts, countdowns, supplies, time blocks, objectives, recipes, planned meals, trips, chores, expiry dates, routines, gift ideas, decisions and list items (also clear_done_tasks, clear_checked_items, remove_list_items, delete_list and clear_meals). One delete or clear is one trash entry; restoring it puts all of its entries back. Not covered: logged entries (use undo_last_log), parts of an entry (a key result, a packing item) and habits, goals, budgets, reminders and templates. empty_trash removes one item, everything older than N days or all of it for good.
  • Recommendation ledger (get_recommendation_history, mark_recommendation): every suggestion get_recommendations gives is filed with when it first appeared and how often. Mark one followed or skipped. For suggestions built on a daily average (sleep, hydration, stress) the ledger later compares that metric in the week before with the days after, with "moving the right way" or "about the same". It is arithmetic on your own logs, not proof of cause and not medical advice. An open suggestion that has not appeared for 30 days closes itself.
  • Pinned backups (pin_backup, unpin_backup): a pinned backup is never rotated away and does not count towards the keep limit (at most 10 pins). get_backup_status lists them. Pin one before a big clean-up.

The doctor and get_repair_suggestions check the trash and the ledger, and search_everything searches the ledger.

Daily Priorities, Plan My Day, Life Wheel, Library & Calendar Export

"my priorities today: finish the report, gym, call mum"    "plan my day"
"I finished the report"                                    "move yesterday's priorities to today"
"rate my life: health 7, career 8, money 3"                "how is my life wheel?"
"add Severance to my shows"  "I finished Dune, 5 stars"    "what should I read next?"
"export my calendar"
  • Priorities: at most three a day, free text or an open task. Completing a priority also completes its task. Hermes nudges you in the afternoon when some are still open.
  • Plan my day: priorities first, then overdue and due tasks, each placed in a free gap of your time blocks using the task's minutes estimate (30 by default). The default is a dry run; apply=true saves time blocks. Tasks already planned are never placed twice.
  • Life wheel: health, career, money, relationships, growth, fun, environment and purpose, each 1 to 10, with bars, the change since the previous check-in and the lowest areas.
  • Library: books, movies, shows, games, podcasts and courses with status (want, doing, done, dropped), progress, a 1 to 5 rating, a suggestion for what to pick up next and yearly stats.
  • Calendar export: an .ics file in your exports folder with time blocks, task due dates, birthdays (repeating yearly), countdowns, trips, expiry dates, chores, renewals and decision reviews. Event ids are stable, so re-importing updates events instead of duplicating them. Import it in Google Calendar (Settings > Import & export), Apple Calendar or Outlook.

All three new stores are included in backup, restore, export, the doctor, search and the recently deleted list.

Inbox & Weekly Review

"inbox: look into cello lessons"        "what's in my inbox?"
"make the cello one a task for Friday"  "drop the lamp idea"
"start my weekly review"                "save my review: wins shipped v1.49, focus on scales and taxes"
  • Inbox: capture first, decide later. process_inbox_item turns an item into a task (with due date, priority and minutes), a note, a library item or a priority, and removes it from the inbox only when the new entry was created.
  • Weekly review: shows the last seven days (finished tasks, priorities done, time blocks, overdue, unsorted inbox, the life wheel's lowest areas) and what you planned to focus on last week, then saves your wins, lessons, rating and next week's focus. create_tasks=true turns the focus items into tasks due next Monday.

Flashcards & Net Worth

"add flashcards: casa = house, perro = dog (deck spanish)"   "quiz me"
"I knew that one"                                           "how are my flashcards doing?"
"my savings account is 4200"  "my loan is 1800 (liability)"  "what is my net worth?"
"show my net worth trend for 12 months"
  • Flashcards: Leitner boxes 1 to 5 with return intervals of 1, 3, 7, 16 and 35 days. Hermes shows the front of the next due card, reveals the back on request and records whether you knew it. Stats cover cards per deck and box, mastered cards and your accuracy.
  • Net worth: assets minus liabilities from the balances you record. Every update is kept, so you get the change since the last update and a month-end trend. Amounts have no currency.

Projects & Waiting For

"start a project: launch the website, due December 1"     "add a task to the website project: write copy"
"how is the website project going?"                       "pause the website project"
"I am waiting for the contract from Sam, check in Friday" "I chased Sam, ask again in 3 days"
"the contract arrived"                                    "what am I waiting for?"
  • Projects: an outcome with a status (active, paused, done), an optional due date and a tag. Tasks carrying the tag count toward progress; Hermes shows the next action and warns about overdue projects and active projects with no open task. Deleting a project keeps its tasks.
  • Waiting for: things you handed to someone else, each with a follow-up date. Due items show in the agenda and as a nudge; chase_waiting records a follow-up and pushes the date out.

Project Structure

graph LR
    A[hermes-life-os] --> B[skills/]
    A --> C[environments/]
    A --> D[demo/]
    A --> E[tests/]
    A --> F[docs/]

    B --> B1[life-os/SKILL.md<br/>Daily rhythm playbook]
    C --> C1[life_os_env.py<br/>Atropos RL environment]
    C --> C2[life_os_config.yaml<br/>Training config]
    D --> D1[demo_life_os.py<br/>CLI / chat / voice orchestration]
    D --> D2[storage.py<br/>Persistence layer]
    D --> D3[patterns.py<br/>Trend detection]
    D --> D4[analytics.py<br/>Pearson correlation engine]
    D --> D5[tools.py<br/>dispatch_tool + TOOLS schema]
    D --> D6[scheduler.py<br/>Cron-style trigger engine]
    D --> D7[notifications.py<br/>console/webhook/Telegram/email]
    D --> D8[run_scheduler.py<br/>Production scheduler entry point]
    D --> D9[plugins.py<br/>Community tool plugin loader]
    D --> D12[life_score.py<br/>Composite 0-100 wellbeing score]
    D --> D13[achievements.py<br/>Streak &amp; milestone badges]
    D --> D14[wrapped.py<br/>Shareable summary card]
    D --> D15[recommendations.py<br/>Rule-based suggestion engine]
    D --> D16[weather.py<br/>Open-Meteo weather correlation]
    D --> D17[life_review.py<br/>Quarterly/yearly retrospective report]
    D --> D18[leaderboard.py<br/>Opt-in household/team leaderboard]
    D --> D19[prompts.py<br/>Deterministic daily reflection prompts]
    D --> D20[moon.py<br/>Local moon phase correlation]
    D --> D21[sleep_debt.py<br/>Sleep debt & bedtime suggestions]
    D --> D22[heatmap.py<br/>GitHub-style SVG contribution heatmap]
    D --> D23[focus_timer.py<br/>Terminal Pomodoro timer]
    D --> D10[users.py<br/>Multi-user registry]
    D --> D11[slack_bot.py<br/>Slack Socket Mode bot]
    E --> E1[test_life_os_env.py]
    E --> E2[test_analytics.py]
    E --> E3[test_storage.py]
    E --> E4[test_tools.py]
    E --> E5[test_scheduler.py]
    E --> E6[test_notifications.py]

    style B1 fill:#27ae60,color:#fff
    style C1 fill:#8e44ad,color:#fff
    style D1 fill:#2980b9,color:#fff
    style D6 fill:#e67e22,color:#fff
    style D7 fill:#e67e22,color:#fff
    style D9 fill:#c0392b,color:#fff
    style D10 fill:#c0392b,color:#fff
    style D11 fill:#c0392b,color:#fff

demo_life_os.py used to be a single ~1600-line file. It's now a thin CLI/chat/voice orchestration layer that imports its storage, pattern detection, and tool-dispatch logic from focused sibling modules - each independently testable and reusable.

Scheduling & Notifications

demo/scheduler.py implements the "Daily Rhythm" cron table from skills/life-os/SKILL.md (07:00 morning, 12:00 midday, 18:00 evening, Monday 08:00 weekly, 20:00 proactive nudge check) as a dependency-free polling loop. The scheduling logic itself (due_entries) is pure and fully unit tested; the actual briefing generation and delivery are injected as callables, so the core engine has no dependency on the OpenAI client or network access. The 20:00 nudge check is LLM-free (see "Proactive Nudges" above) and stays silent when there's nothing worth flagging.

Since v1.37.0 the scheduler loop also accepts a dynamic_source hook for things that are not a fixed (time, mode) pair. run_scheduler.py plugs in reminders.scheduler_source, so every timed reminder you create with create_reminder is delivered through the same notification channel at its saved time (see "Reminders" below), at most once per day, and a failing source or notifier never stops the loop.

demo/notifications.py delivers briefings through a pluggable channel, selected via HERMES_NOTIFY_CHANNEL: console (default), webhook, telegram, or email (SMTP). All channels are stdlib-only. A failed remote channel never crashes the scheduler - it's caught, logged, and the briefing still prints to console.

To run the scheduler in production:

set ANTHROPIC_API_KEY=sk-ant-...   # or OPENAI_API_KEY / OPENROUTER_API_KEY / a running ollama
set HERMES_NOTIFY_CHANNEL=telegram
set TELEGRAM_BOT_TOKEN=...
set TELEGRAM_CHAT_ID=...
python demo/run_scheduler.py

Voice Mode

python demo/demo_life_os.py --voice
# or pin a backend/model explicitly:
python demo/demo_life_os.py --voice --provider anthropic --model claude-sonnet-5

Speak to Hermes directly. It listens via microphone, processes your input using everything it knows about you, and responds out loud via system TTS.

No extra API key needed - uses built-in Windows/Linux speech synthesis.

To stop: say or type exit


Dashboard

pip install "hermes-life-os[dashboard]"   # or: pip install matplotlib (running from source)
hermes-life-os-dashboard
hermes-life-os-dashboard --days 60 --compare-days 7 --out my-report.html

Turns your logged mood/sleep/stress/energy/hydration data and the correlations Hermes already detects (e.g. "poor sleep tracks with lower mood, r=0.62") into a single self-contained HTML report with charts - opens straight in your browser, no server, nothing leaves your machine. Needs no LLM/API key at all - it's pure local data analysis. Includes a retrospective section comparing this week to last week (--compare-days changes the window size), color-coded by whether the change is favorable - a stress increase shows red, a mood increase shows green.

Example dashboard trend chart

Example output from 28 days of sample data - your own chart will reflect whatever you've actually logged.

Live Web Dashboard

pip install "hermes-life-os[web]"   # or: pip install flask
hermes-life-os-web
# open http://127.0.0.1:8080

The interactive, always-current counterpart to the static HTML report above - same trends/correlations/retrospective/habit data, served as JSON and rendered client-side with Chart.js, so switching the day-range re-fetches and re-draws instantly instead of regenerating a file. Localhost-only by default and read-only (no API key needed, unlike the Local REST API below) - it only ever reads your own data for your own browser. See demo/web_dashboard.py's docstring before changing --host beyond 127.0.0.1.

Goal Tracking

Goals can track themselves from real data instead of needing manual progress updates - just tell Hermes what to track:

  • "Set a goal to sleep 7+ hours a night" -> auto-tracks against your logged sleep, direction "at_least"
  • "Set a goal to keep stress under 4" -> direction "at_most"
  • "How am I doing on my goals?" -> recomputes and reports current progress

Progress is the average of the linked metric over a rolling window (7 days by default) relative to the target, clamped to 0-100%. Goals without a linked metric keep working exactly as before - a plain percentage you update manually.

Health Data Import

pip install hermes-life-os
hermes-life-os-import --apple-health export.xml
hermes-life-os-import --csv my_data.csv
hermes-life-os-import --csv my_data.csv --dry-run   # preview without writing

Reduces manual one-entry-at-a-time logging by bulk-importing data you already have, with real historical dates preserved (not stamped "today"):

  • Apple Health (export.xml from Health app -> profile icon -> Export All Health Data): imports Sleep Analysis and Dietary Water records, aggregated per day.
  • Generic CSV: any file with a date column (YYYY-MM-DD) plus any subset of sleep_hours, mood, stress, energy, hydration columns - works for a Google Fit CSV export or your own spreadsheet.

Imported entries are tagged so they're distinguishable from entries logged live through chat.

Calendar Import (meeting load vs. mood/stress)

hermes-life-os-calendar --ics calendar.ics
hermes-life-os-calendar --ics calendar.ics --dry-run

Correlates meeting-heavy days with mood/stress/sleep - no OAuth or live API needed, just a standard .ics export (Google Calendar: Settings -> Import & export -> Export; Outlook: File -> Save Calendar; Apple Calendar: File -> Export). Only timed events count toward meeting hours (all-day events are skipped); recurring events count once, on their start date. Once imported, ask Hermes "is my stress linked to meeting-heavy days?" or check the dashboard's Correlations section.

Deeper Analysis: Anomalies & Before/After Comparisons

  • "Has anything been unusual lately?" -> flags statistical outlier days (e.g. "today's stress was far above your normal range")
  • "Did starting meditation on March 1st actually help?" -> compares metric averages before vs. after a specific date, instead of just a fixed weekly window
  • "How was I in March?" / "Summarize last month" -> pulls real averages and notable entries (gratitude, dreams, notes) for any date range you ask about

Proactive Nudges

The scheduler (see above) includes a daily check (20:00 by default) that looks for anything worth flagging - an unusual day, a goal falling behind, a monthly budget that is over or nearly used up - using the same deterministic analysis as the tools above, with no LLM call needed. It stays silent on days with nothing notable, so it won't spam you.

Data Export

hermes-life-os-export --json backup.json --csv summary.csv

Your data isn't locked in. --json writes a complete backup (every memory entry plus every data store - profile, habits, goals, all the logs, spending, budgets, reminders, templates and so on - unmodified, as readable JSON even when encryption is on, since you asked for a copy to take elsewhere). --csv writes a daily summary in the same shape hermes-life-os-import --csv expects - export, edit in a spreadsheet, and re-import elsewhere if you want.

Telegram Bot

set TELEGRAM_BOT_TOKEN=...     # from @BotFather
set TELEGRAM_CHAT_ID=...       # your own numeric chat id
hermes-life-os-telegram

Talk to Hermes from your phone - no server, no webhook, just long polling (keep the process running, e.g. in tmux/screen or as a background service). Only messages from TELEGRAM_CHAT_ID are ever processed, so your data stays private even if someone finds your bot's username. See demo/telegram_bot.py's docstring for the exact setup steps (getting a token and finding your chat id).

Model reliability note: this project defines a lot of tools (27+). Small/CPU-friendly local models (e.g. llama3.2:3b) can struggle to use them reliably - they may log things you didn't ask for, or occasionally emit a raw tool-call attempt as plain text instead of actually calling the tool (Hermes detects and filters that specific failure so you never see raw JSON, but the underlying action still won't happen). llama3.1 (8B) is noticeably more reliable via Ollama, at the cost of being slower on CPU-only machines (several minutes per reply). Any cloud provider (Anthropic/OpenAI/OpenRouter) is both faster and more reliable if you have API access.

Discord Bot

pip install "hermes-life-os[discord]"   # or: pip install discord.py
set DISCORD_BOT_TOKEN=...      # from the Discord Developer Portal
set DISCORD_USER_ID=...        # your own numeric Discord user id
hermes-life-os-discord

The Discord counterpart to the Telegram bot above - same idea (talk to Hermes, log meals from a photo, send a voice message), different platform. Uses discord.py's own event-driven client under the hood (rather than the Telegram bot's hand-rolled long-polling loop), so it works in a DM or in any server channel the bot can see. Only messages from DISCORD_USER_ID are ever processed - everyone else is silently ignored. See demo/discord_bot.py's docstring for the exact setup steps (creating a bot application, enabling the Message Content intent, and finding your user id). The same vision-model requirement for photo meal logging applies here - see the Photo Meal Logging section below.

WhatsApp Bot

pip install "hermes-life-os[whatsapp]"   # or: pip install flask twilio
set TWILIO_ACCOUNT_SID=...
set TWILIO_AUTH_TOKEN=...
set WHATSAPP_ALLOWED_NUMBER=whatsapp:+1XXXXXXXXXX   # your own number, E.164
hermes-life-os-whatsapp

The third chat platform, via Twilio's WhatsApp API. Unlike the Telegram bot (polling) and Discord bot (websocket client), this is a webhook server - Twilio pushes messages to it, so it needs to be reachable from the internet (ngrok http 8766 works well for personal use with Twilio's free WhatsApp Sandbox). Same feature set as the other two: plain text, photo-based meal logging, and voice-note transcription. Every incoming request's Twilio signature is verified before anything is processed, on top of the same single-number authorization the other bots use. Full setup steps (sandbox join code, webhook URL) are in demo/whatsapp_bot.py's docstring. The same vision-model requirement for photo meal logging applies here too - see the Photo Meal Logging section below.

Slack Bot

pip install "hermes-life-os[slack]"   # or: pip install slack_bolt
python demo/users.py add alex --profile alex
python demo/users.py link alex slack U0123ABC   # your Slack member ID
set SLACK_BOT_TOKEN=xoxb-...
set SLACK_APP_TOKEN=xapp-...
hermes-life-os-slack

The fourth chat platform, and the first one built multi-user from the start: DM the bot and it's automatically routed to your profile, so a whole household or team can share one bot process. Uses Socket Mode (a persistent websocket, via Slack's own slack_bolt framework) - same "no public server, no webhook" philosophy as the Telegram bot, just using Slack's officially supported connection handling instead of a hand-rolled polling loop. Works single-user too (SLACK_ALLOWED_USER_ID, same pattern as the Discord/Telegram bots) if you don't need the multi-user registry. Same photo-based meal logging support as the other bots. Full setup (creating the Slack app, required scopes, finding your member ID) in demo/slack_bot.py's docstring and docs/MULTI_USER.md.

Local REST API

pip install "hermes-life-os[api]"   # or: pip install flask
set LIFE_OS_API_KEY=some-long-random-string
hermes-life-os-api

A lightweight, localhost-only HTTP API for third-party integrations that don't want to (or can't) go through an LLM at all - Apple Shortcuts, Android Tasker, a browser extension, a home-screen widget, an Alfred/Raycast workflow, curl in a cron job, etc. Exposes the same tools the chat agent uses:

curl -H "X-API-Key: some-long-random-string" http://127.0.0.1:8765/api/tools
curl -H "X-API-Key: some-long-random-string" -X POST \
     -d '{"score": 8}' http://127.0.0.1:8765/api/tools/log_mood

Binds to 127.0.0.1 by default and refuses to start without LIFE_OS_API_KEY set - every request needs it as an X-API-Key header. See demo/local_api.py's docstring for the full endpoint list and the security notes on exposing this beyond your own machine. For a step-by-step Apple Shortcuts / Android / browser bookmarklet setup, see docs/SHORTCUTS.md.

Weather Correlation

"does the weather affect my mood?"
"any connection between rain and my energy levels?"

demo/weather.py fetches historical daily weather (temperature, precipitation) for a place you name, via Open-Meteo's free, keyless API (https://open-meteo.com) - no signup, no API key, no cost - and correlates it against your tracked metrics the same way the core correlation engine does. This is the only tracker that makes a network call (every other one is 100% local), and it only ever does so when explicitly asked, sending nothing but the place name you provide.

Life Review - the flagship retrospective report

hermes-life-os-review --days 90              # a quarter -> hermes-life-review.html
hermes-life-os-review --days 365 --out my-year-review.html

The "big" report - ties together everything else Hermes tracks into one self-contained HTML page: your Life Score trend chart, best/toughest day, a period-over-period retrospective (this quarter vs. the one before), every correlation detected (same-day and lagged), achievements earned, and habit streaks. Same "no server, no JS build step" approach as the dashboard, just built for a much longer lookback window - a quarter or a year, rather than 30 days. Doesn't include weather correlation automatically, since that's the one feature that touches the network - ask for it separately if wanted.

Need something printable or shareable outside a browser? --format pdf renders a two-page PDF version with the same data (Life Score, retrospective, correlations, achievements, habits) on a light, print-friendly background:

hermes-life-os-review --days 90 --format pdf --out my-quarter.pdf

hermes-life-os-wrapped supports the same trick even more simply - just name the output file .pdf instead of .png and it's picked up automatically:

hermes-life-os-wrapped --out hermes-wrapped.pdf

Leaderboard - opt-in household/team comparison

"join the leaderboard"
"show me the leaderboard"

Builds on the multi-user system (see docs/MULTI_USER.md): a friendly, opt-in ranking across every profile on the install, by average Life Score, current logging streak, and achievements earned. Nobody is included by default - a profile only appears after explicitly opting in (join_leaderboard, or python demo/leaderboard.py join), and leaving again takes effect immediately. Only those three numbers are ever shared across profiles - no journal content, no raw logged entries, and nothing ever leaves the machine.

Streak Freezes

Habits (update_habit) now bank a streak freeze every 7 days of an active streak, up to 3 at once - a small forgiveness mechanic so one missed day doesn't erase weeks of consistency:

"I meditated" (7 days running) -> "Habit 'meditate': streak 7 days (best: 7) - earned a streak freeze! (1 available)"
"I missed meditating today, use a freeze" -> "Habit 'meditate': streak protected with a freeze! Still at 7 days (best: 7). 0 freeze(s) left."

No freeze banked and a day's missed? The streak resets to 0, same as always - freezes are a bonus for consistency, not a way around ever having an off day.

On This Day & Daily Prompts

"what was I doing on this day last year?"
"give me a reflection prompt"

Two small additions aimed at journaling depth, not just numbers:

  • On This Day (get_on_this_day) - a nostalgia lookup: finds memory entries logged on today's month/day in previous years, most recent first.
  • Daily Prompt (get_daily_prompt, demo/prompts.py) - a rotating reflection question, deterministic by calendar date (same day always returns the same prompt, and it changes daily) - no state to persist, no randomness to make testing flaky.

Contribution Heatmap

hermes-life-os-heatmap --days 365 --out heatmap.svg
hermes-life-os-heatmap --days 90 --out heatmap.html    # wrapped with stats

A GitHub-style calendar heatmap of your logging activity - darker squares for more active days, same five-level palette as GitHub's own contribution graph. Pure SVG (no matplotlib needed), so it's fast and embeddable anywhere that accepts raw SVG or an <img> tag. .html output wraps it with active-day/streak stats above the grid.

Moon Phase Correlation

"does the full moon affect my mood?"

demo/moon.py correlates lunar phase against tracked metrics, purely via local astronomical calculation - unlike weather.py, this makes zero network calls; the phase is computed on-device from a known reference new moon date. Offered in the same evidence-based spirit as every correlation Hermes computes: not because lunar effects on mood are scientifically established (most rigorous research says they aren't), but because it's a fun, harmless pattern to check against your own data - if Hermes finds nothing, that itself is the honest answer.

Sleep Debt & Suggested Bedtime

"how much sleep debt do I have?"
"what time should I go to bed if I'm waking up at 7?"

demo/sleep_debt.py sums the shortfall between a target sleep duration (default 8h) and what's actually logged over a recent window - only days with a logged sleep value count, so gaps in logging are never mistaken for debt. get_suggested_bedtime factors existing debt into tonight's recommended bedtime, nudged earlier gradually (a quarter of the debt per night, capped at 1h extra) rather than proposing something unrealistic in a single night.

Day-of-Week Insights

"what's my best day of the week?"
"is my mood worse on Mondays?"

demo/day_of_week.py buckets already-logged mood, energy, stress, sleep, hydration, meeting-hours, and readiness by weekday, then reports which day averages best and worst for each metric - lower is treated as "best" for stress specifically, since a calmer day is the good outcome there. Purely local arithmetic over daily_averages(), no new tracking of its own; a metric needs at least two distinct weekdays logged before Hermes will call one "best" and another "worst".

Habit Milestones

"how close am I to a milestone on my habits?"

demo/habit_milestones.py looks at every habit with an active streak and counts down to its next round-number milestone (7, 14, 30, 50, 100, 150, 200, 365, 500, 1000 days, doubling beyond that). This is a different lens than the near-milestone nudge already surfaced opportunistically by recommendations.py - get_habit_milestones gives the full countdown list for every active habit on demand, rather than waiting for one habit to get close enough to be worth a nudge.

Logging Consistency

"how consistent have I been with logging?"
"which trackers have I been neglecting?"

demo/consistency.py (get_logging_consistency) reports the percentage of recent days that had at least one entry logged, overall and broken down per tracked metric. A metric that's never been logged in the window is simply left out of the breakdown rather than shown as a discouraging 0% - an unused tracker isn't "inconsistent", it's just unused.

Time-of-Day Insights

"am I happier in the mornings?"
"is my stress worse at night?"

demo/time_of_day.py (get_time_of_day_insights) buckets individual logged entries into Morning/Afternoon/Evening/Night by the hour they were logged and reports which time of day is best and worst per metric - a companion to Day-of-Week Insights that looks at time-of-day rather than day-of-week, and works entry-by-entry rather than one daily average, since mood at 8am and mood at 8pm on the same day should land in different buckets.

Month-over-Month Comparison

"how is this month going compared to last month?"

demo/monthly_summary.py (get_monthly_comparison) compares this calendar month so far against the same day-count span at the start of last calendar month - e.g. running it on the 10th compares days 1-10 of each month, not the whole of last month against a partial current one. Handles month-length edge cases (a 31-day January comparing against a 28/29-day February) by capping the previous-month span at that month's own last day. Built on top of the existing compare_periods() averaging/delta logic, just with calendar-month date ranges instead of rolling day windows.

Habit Personal Bests

"am I close to beating my meditation streak record?"

demo/habit_pb.py (get_habit_pb_progress) compares every habit's current streak against its own personal-best streak - a new best, a tie, or how many days remain to tie it. This is a different, often more motivating lens than the fixed round-number countdown in Habit Milestones: a habit whose best streak is 12 will never feel "close" to milestone 14, but "2 days from tying your best" always does.

Workout Summary

"give me a fitness recap"
"how's my workout streak?"

demo/workout_summary.py (get_workout_summary) totals recent workouts, breaks them down by type, and reports the current consecutive-day workout streak. The streak always looks at your full workout history rather than just the requested lookback window, so a streak that started 40 days ago still shows correctly even when asking for a 30-day summary.

Meditation Summary

"how's my meditation practice going?"

demo/meditation_summary.py (get_meditation_summary) totals recent meditation sessions and reports what percentage of days in the window had at least one session - multiple sessions on the same day count once toward that consistency figure.

Gratitude Recap

"what have I been grateful for lately?"

demo/gratitude_recap.py (get_gratitude_recap) totals recent gratitude entries, surfaces the most recurring words across them as a lightweight signal for recurring themes (not real NLP, just frequency counting with a small stopword list), and lists the most recent items.

Meal Summary

"give me a nutrition recap"

demo/meal_summary.py (get_meal_summary) totals recent meals, reports average calories per meal and per logged day, breaks meals down by meal time (breakfast/lunch/dinner/snack), and lists the most frequently logged foods (case-insensitively, so "Salad" and "salad" count together).

Hydration Summary

"how well have I been hitting my water goal?"

demo/hydration_summary.py (get_hydration_summary) rebuilds daily totals from the raw hydration log (rather than the single "today" counter, which only reflects the current calendar day), then reports the average glasses/day against your goal, how many logged days hit it, and the current consecutive-day goal-met streak.

Focus Summary

"give me a productivity recap"

demo/focus_summary.py (get_focus_summary) totals recent focus sessions and reports total/average duration, average quality, completion rate, total distractions, and the most common task.

Dream Recap

"what are my recurring dream themes?"

demo/dream_recap.py (get_dream_recap) totals recent dreams and surfaces which symbols or emotions recur across multiple distinct dreams (a symbol mentioned twice in the same dream only counts once, so recurrence reflects a genuine pattern across nights rather than repetition within one dream). Dreams have no dedicated storage file of their own - they live only in the shared memory log - so this reads from get_recent_memory rather than a load_*() function like the other summary tools.

Stress Summary

"give me a stress recap"
"is my stress getting better or worse?"

demo/stress_summary.py (get_stress_summary) reports the average stress score, a first-half-vs-second-half trend within the window ("improving", "worsening", or "steady" - a difference under 1 point on the 0-10 scale is treated as steady, to avoid flip-flopping on small amounts of data), how many days crossed a high-stress threshold, and the most common triggers.

Habit Consistency, Mood Impact & Overview

"how consistent have I really been with meditation?"
"does meditating actually improve my mood?"
"how are my habits doing overall?"

update_habit now also records every check-in (done, missed, or protected with a freeze) as a habit_completion entry in the shared memory log, alongside the existing streak counters in habits.json. This unlocks three new views built on that real per-check-in history rather than just the current streak state:

  • demo/habit_consistency.py (get_habit_consistency) - the percentage of check-ins that were successful for each habit, which can tell a very different story than the current streak (a short current streak can follow a long run of consistency, or vice versa).
  • demo/habit_correlation.py (get_habit_mood_impact) - compares average mood on days a given habit was completed versus days it was missed. This is naturally forward-looking: it only reflects habit check-ins recorded after this feature shipped, and needs mood logged on both completed and missed days to say anything meaningful.
  • demo/habit_overview.py (get_habit_overview) - a single combined dashboard over Habit Milestones, Habit Personal Bests, and Habit Consistency, for a general "how are my habits doing" question instead of picking one specific angle. Reuses each module's own formatter, so the wording matches calling that tool directly.

Reading Pace

"how fast am I reading Atomic Habits?"
"log_reading('War and Peace', pages=40, total_pages=1225)"

log_reading now accepts an optional total_pages for a book. demo/reading_pace.py (get_reading_pace) computes pages/day for a specific title and, once a total page count has ever been logged for it, estimates the remaining days and a finish date at the current pace. Without a total_pages on file it still reports the pace, just without an estimate - there's nothing to count down to.

Spending Trends

"which spending categories are trending up?"

demo/spending_trends.py (get_spending_trends) compares spending per category over a recent window against the equal-length window before it - a different angle than the existing Spending Summary, which only totals a single window. A category that's new or has stopped appearing still shows up, with 0 on the missing side, rather than silently disappearing from the comparison.

Substance-Sleep Impact

"does caffeine affect my sleep?"

demo/substance_correlation.py (get_substance_sleep_impact) compares average sleep hours on days a given substance was logged versus days it wasn't. This compares same-calendar-day entries rather than "the following night", since a sleep log's date is whatever day it was logged (typically the morning after waking) and there's no reliable way to know which night it refers to otherwise.

Social Insights & Mood Impact

"who have I been spending time with lately?"
"does socializing actually improve my mood?"

demo/social_insights.py (get_social_insights) breaks recent social interactions down by who they were with - time spent, average quality, and the most frequent contact - a different angle than the existing Social Summary, which only totals the whole window. demo/social_correlation.py (get_social_mood_impact) compares average mood on days with a logged social interaction versus days without one, using the same same-calendar-day comparison convention as Habit Mood Impact and Substance-Sleep Impact.

Medication Streak

"how consistent have I been with taking my Vitamin D?"

demo/medication_streak.py (get_medication_streak) reports the current and longest consecutive-day streak of taking a specific medication or supplement - a different angle than the existing Medication Adherence, which only reports an overall percentage for the lookback window rather than a day-by-day streak.

Workout Mood Impact

"does exercising actually improve my mood?"

demo/workout_correlation.py (get_workout_mood_impact) compares average mood on days a workout was logged versus days without one, using the same same-calendar-day comparison convention as Habit Mood Impact, Substance-Sleep Impact, and Social Mood Impact.

Life Review, Now Conversational (and Two More Matplotlib Crash Fixes)

"give me a fuller review of the last quarter"

Fixing wrapped.py's matplotlib crash last round turned up the exact same bug in two more report generators: demo/dashboard.py and demo/life_review.py both imported matplotlib at module level and called sys.exit(1) if it wasn't installed - which would have killed the whole bot/API process the first time a text-only caller touched either module in an environment without matplotlib. Both now use a shared _get_pyplot() lazy-import helper (dashboard._get_pyplot(), reused by life_review.py) that raises a normal ImportError only when a chart is actually about to be rendered; build_dashboard_data() and build_life_review_data() need no plotting library at all.

That fix made demo/life_review.py's data cheap to expose conversationally too: new format_life_review_summary() gives a text version - life score trend, best/toughest day, correlation insights, a retrospective comparison against the prior period, badges earned - available as get_life_review, alongside the existing hermes-life-os-review HTML/PDF CLI. It's the deeper counterpart to get_wrapped's quick recap.

Proactive Nudges & Wrapped, Now Conversational

"is there anything I should know about?"
"give me my wrapped for the last 30 days"

Two more existing-but-CLI/scheduler-only features are now ordinary tools: demo/nudges.py's deterministic anomaly and off-track-goal checker (previously only run by the scheduler for proactive notifications) is now available on demand as get_nudges, and demo/wrapped.py's "Spotify Wrapped"-style period recap - average life score, mood, sleep, stress, best day, badges earned - is now available as text via get_wrapped, alongside its existing hermes-life-os-wrapped image-card CLI.

Getting get_wrapped working safely surfaced a real bug: wrapped.py imported matplotlib at module level and called sys.exit(1) if it wasn't installed - fine for a CLI script, but it would have killed the entire bot/API process the first time a text-only caller tried to use it in an environment without matplotlib. The matplotlib import is now lazy, living inside render_wrapped_image() (the only function that actually needs it) and raising a normal ImportError instead of exiting the process; build_wrapped_stats() and the new format_wrapped_summary() need no plotting library at all.

Correlation Refactor, Reading Mood Impact & Insights Digest

"what stands out in my data lately?"
"does reading affect my mood?"

Habit Mood Impact, Substance-Sleep Impact, Social Mood Impact, and Workout Mood Impact had each grown their own near-identical "average metric X on days condition Y held versus days it didn't" implementation. demo/correlation_utils.py now holds that logic once

  • compute_metric_impact_by_presence() for checks where "it didn't happen" is inferred from absence (workouts, substances, social interactions, reading), and compute_metric_impact_by_status() for checks where "it didn't happen" needs its own explicit record (a habit logged as missed, as opposed to a habit that was never checked in on at all). Every existing correlation tool now calls one of these two functions internally; their public functions, return shapes, and wording are unchanged.

That shared foundation made two things cheap to add:

  • Reading Mood Impact (demo/reading_correlation.py, get_reading_mood_impact) - the same day-vs-day mood comparison, now for reading sessions.
  • Insights Digest (demo/insights_digest.py, get_insights_digest) - runs every correlation check at once (workout/social/reading vs mood, every tracked habit vs mood, every logged substance vs sleep) and surfaces only the findings with a meaningful difference, instead of the person needing to know which of seven-plus separate tools to call, or call all of them just to see if anything stands out. A digest that repeats "not enough data" seven times isn't useful, so anything below the threshold (default 1.0 points/hours) is silently omitted.

Export & Backup, Now Conversational

"export my data as csv"
"back up my data right now"

Hermes Life OS has always had a full data-export system (demo/data_export.py - JSON/CSV/Markdown) and rotating automated backups (demo/backup.py), but until now they could only be triggered by running those scripts directly. demo/export_tool.py wraps them as two ordinary tools, export_data and backup_now, so they can be triggered from a normal conversation - picking a sensible default output path under the active profile's exports/ or backups/ directory so nobody has to supply one. export_data is for taking your data elsewhere (a spreadsheet, Obsidian, Notion); backup_now is for safekeeping in place.

Habit-Linked Goals

"update_goal('meditation streak', linked_habit='meditate', target_streak=30)"

update_goal now supports a third way to track progress, alongside manual and metric-linked: linked_habit + target_streak ties a goal's progress directly to a habit's actual current streak (e.g. "hit a 30-day meditation streak"), computed via a new compute_habit_goal_progress() in analytics.py - the habit-tracking counterpart to the existing metric-tracking compute_goal_progress(). check_goal_progress refreshes habit-linked goals from the live streak the same way it already refreshes metric-linked ones. A goal is manual, metric-linked, or habit-linked - not more than one at once.

Quick-Log Templates

"save my usual breakfast as a template"
"log my usual breakfast"

demo/templates.py lets you save a named shortcut for a specific log_* tool call with fixed parameters (save_log_template), then replay it later with one call (use_log_template) instead of re-typing the same details every time - list_log_templates and delete_log_template round it out. Templates can only ever point at log_* tools, never at export/backup/goal/habit-management or any other tool, so replaying one can only ever add a new log entry, never trigger a side effect beyond what the person explicitly asked to save.

List, Delete Habits & Goals

"what habits do I have set up?"
"delete my old 'read more' goal"

A real gap: a brand-new habit with a 0-day streak (created via a missed first check-in, or one that's since dropped to a 0-day streak with no past best) was invisible everywhere - get_habit_milestones and get_habit_pb_progress both explicitly exclude habits with no active or past streak, and the health dashboard only lists active ones. list_habits now shows every habit regardless of streak status. delete_habit and delete_goal round out basic CRUD for both - before this, there was no way to remove a habit or goal at all once created, only delete_entry for individual memory entries.

Budgets

"set a 400/month budget for groceries"
"how am I doing on my budgets this month?"

demo/budgets.py is the project's first forward-looking money tool - every existing spending feature (log_expense, get_spending_summary, get_spending_trends) is retrospective, reporting what already happened, with no concept of a limit. set_budget saves a per-category monthly limit; get_budget_status compares it against actual spending logged so far in the current calendar month (reusing load_spending, no new tracking of its own) and flags any category that's over. list_budgets shows the raw limits, and delete_budget removes one. Setting a budget for a category that already has one overwrites the limit rather than creating a duplicate.

"will I stay within my budgets this month?"
"how did I do on my budgets over the last 3 months?"
"where is my money going outside my budgets?"

Three more tools build on the same data (v1.37.0):

  • get_budget_forecast projects month-end spending from this month's pace so far and says how much you can still spend per remaining day. Each budget is over, will_exceed or on_track, and projections made in the first days of the month are flagged as rough.
  • get_budget_history looks back over the last N completed months (default 3, max 24) and shows spending per month and how many months went over. Budgets are not versioned, so every month is measured against the current limit.
  • get_unbudgeted_spending lists this month's spending in categories that have no budget at all, biggest first.

Spending categories now match budgets case-insensitively, so an expense logged as "Food" counts against a "food" budget (before, it was silently missed). Budgets that are over, or at least 80% used, also show up in the proactive nudge check and get_nudges.

v1.38.0 adds two more things:

"set a 400 budget for groceries and warn me at 90%"
"what should my budgets be?"
  • set_budget takes an optional alert_pct (1-100) so a budget can warn earlier or later than the default 80% (0 removes a custom value), and it now rejects a negative, non-numeric or infinite limit instead of storing it (a numeric string such as "400" is accepted and converted).
  • suggest_budgets proposes a monthly limit for every spending category that has no budget yet: the average of the last few completed months (default 3, months without spending count as zero) plus a 10% buffer (buffer_pct), rounded up to a tidy number. It only suggests; use set_budget to adopt one.

Reminders

"remind me to stretch every day at 9am"
"remind me to call mom on sundays"
"what reminders do I have set?"

demo/reminders.py gives user-defined reminders a real, queryable home for the first time - scheduler.py's schedule was entirely hardcoded (morning/checkin/evening/weekly/nudge_check/backup), with no mechanism for a person to define their own. create_reminder saves free-text reminders with an optional time hint and optional days of the week; list_reminders and delete_reminder round out CRUD.

As of v1.37.0 reminders are actually delivered. Times are normalized to 24-hour HH:MM (9am, 9:30pm and 17:30 all work; a bare 9 is ambiguous and rejected), and while hermes-life-os-scheduler is running each timed reminder is sent through your notification channel at its saved time on the days it applies to, once per day. A two-minute grace window means a slow poll cannot skip one. A reminder with no time is a day-level note, and one whose time cannot be understood is saved but never fires (the agent tells you so when you create it).

"change my stretch reminder to 6:30pm"
"pause my stretch reminder while I'm away"
"what reminders do I have today?"
  • update_reminder changes the text, time and/or days of an existing reminder (time="" removes the time, days=[] means every day).
  • pause_reminder / resume_reminder switch a reminder off and on without deleting it. Paused reminders are marked in list_reminders.
  • get_todays_reminders lists what applies today in time order.

v1.38.0 adds one-off and short-term reminders:

"remind me on 2026-11-02 at 9am to renew my passport"
"remind me in 20 minutes to check the oven"
"snooze that reminder for 15 minutes"
"clear my old one-off reminders"
  • create_reminder and update_reminder accept an optional date (YYYY-MM-DD), making a one-off reminder that applies on that date only and ignores days.
  • remind_me_in sets a one-off reminder N minutes from now (1 to 1440).
  • snooze_reminder delays any reminder so it fires again in N minutes (default 10); its normal time is suppressed until then.
  • clear_past_reminders deletes one-offs whose date has passed. Recurring and upcoming reminders are never touched.

Goal Deadlines

"update_goal('finish the report', deadline='2026-03-01')"
"what goal deadlines am I about to miss?"

update_goal now accepts an optional deadline (YYYY-MM-DD). demo/goal_deadlines.py (get_goal_deadlines) lists every deadline-bearing goal sorted by urgency, with overdue goals surfaced first and how many days overdue they are - so a goal with a deadline never silently falls off the radar the way a plain progress percentage can.

Focus Timer

hermes-life-os-focus 25 --task "writing"
hermes-life-os-focus 25 --task "deep work" --break 5

A terminal Pomodoro-style countdown that logs itself on completion, via the exact same log_focus_session path a chat-based request uses - so this isn't a separate timer app you also have to remember to log afterward. Ctrl+C during the countdown stops it early and does not log a session (an abandoned session isn't a completed one).

Markdown / Obsidian / Notion Export

hermes-life-os-export --markdown ./obsidian-vault/hermes
hermes-life-os-export --markdown ./notes --days 90

Adds a third export format alongside JSON and CSV: one file per day (YYYY-MM-DD.md) with YAML frontmatter for that day's numeric metrics and a bulleted list of everything logged - filename convention matches Obsidian's Daily Notes plugin exactly, and Notion's markdown importer reads YAML frontmatter as page properties, so the same export folder drops into either tool with no conversion step.

Semantic Memory Search

recall searches by exact keyword; ask Hermes something like "have I felt this way before?" or "find entries about feeling overwhelmed" (even if you never used that exact word) and it can fall back to semantic_recall - a local, free embedding search via Ollama (ollama pull nomic-embed-text first) or OpenAI:

set EMBEDDING_PROVIDER=ollama   # or openai; auto-detects from OPENAI_API_KEY otherwise

Embeddings are cached per entry and only recomputed when that entry's text actually changes.

Oura Ring Import

set OURA_PERSONAL_ACCESS_TOKEN=...   # https://cloud.ouraring.com/personal-access-tokens
hermes-life-os-oura --days 30

No OAuth flow - just a personal access token from Oura's own dashboard. Imports real sleep duration (merges directly with manually logged and Apple-Health-imported sleep) and your daily readiness score (a new tracked metric - ask "is my readiness linked to sleep or stress?").

Weekly Email Summary

set HERMES_SMTP_HOST=smtp.gmail.com
set HERMES_SMTP_PORT=587
set [email protected]
set HERMES_SMTP_PASSWORD=...          # an app password, not your real password
set [email protected]      # optional, defaults to HERMES_SMTP_USER

hermes-life-os-weekly-email
hermes-life-os-weekly-email --days 30 --compare-days 7

Emails the same self-contained report hermes-life-os-dashboard generates - trend charts, correlations, retrospective, habit streaks - straight to your inbox. Not wired into the scheduler by default (not everyone has SMTP configured); schedule it yourself with your OS's own task scheduler if you want it automatic weekly:

Linux/macOS (cron):        0 8 * * 1  hermes-life-os-weekly-email
Windows (Task Scheduler):  weekly trigger, action = same command

Voice Notes (Telegram)

pip install "hermes-life-os[voice]"   # or: pip install faster-whisper

Send the Telegram bot a voice note instead of typing - it's downloaded and transcribed locally (a free Whisper model via faster-whisper, no cloud API, no per-minute cost) before being processed exactly like a typed message. The reply is prefixed with what Hermes heard, so you can catch a bad transcription. WHISPER_MODEL (default base) controls speed vs. accuracy - tiny is fastest, small/medium are more accurate but slower on CPU-only machines. Without faster-whisper installed, voice notes get a clear "couldn't process" reply instead of silently failing.

Photo Meal Logging (Telegram, Discord, WhatsApp & Slack)

Send any of the four bots a photo of your meal (with an optional caption) and a vision-capable LLM identifies what's in it and logs it

  • no separate step needed. OpenAI's and Anthropic's default models already support vision. On Ollama, pull a vision-capable model yourself (ollama pull llava) and point Hermes at it explicitly (set HERMES_MODEL=llava or --model llava) - Ollama's default text-only models (like llama3.1) will simply ignore the image.

Automatic Backups

Hermes takes a timestamped local backup of your data every day at 20:30 (right after the evening nudge check), keeping the 7 most recent by default and pruning older ones. Backups live alongside your other data (<profile dir>/backups/). Run it manually anytime:

hermes-life-os-backup            # keep the default 7
hermes-life-os-backup --keep 14  # keep the 14 most recent

A backup holds the whole memory journal plus every data store registered in storage.DATA_STORES (18 of them), and a small _meta block with the format version. Before v1.38.0 only nine stores were included, so the nightly backup silently left out spending, reminders, budgets and more. Older backups still restore fine; they just lack those sections.

If LIFE_OS_ENCRYPTION_KEY is set, backups are encrypted with it, so an encrypted profile no longer leaves a plaintext copy of itself in backups/. Keep the passphrase: an encrypted backup cannot be read without it.

Through chat, backup_now takes a backup on demand and get_backup_status reports how many exist, when the newest was taken, its size, whether it is complete and whether it is encrypted. (backup_now always keeps at least one backup; a keep of 0 or a nonsense value from the agent used to be able to delete the backup it had just written.) If backups exist but the newest is 3 or more days old, or cannot be read, the proactive nudges say so.

Restore

hermes-life-os-restore --list                      # backups of this profile, newest first
hermes-life-os-restore --latest --dry-run          # show exactly what would change
hermes-life-os-restore --latest                    # asks for confirmation first
hermes-life-os-restore backup-2026-10-01-203000.json --only spending,budgets --yes
hermes-life-os-restore backup.json --profile alex

Until v1.38.0 backups could be written but never read back. Restoring is deliberately a command-line action, not a chat tool. It is built to be hard to get wrong:

  • The whole file is validated before anything is written - every section must have the right shape and every --only name must exist. One problem and nothing is touched.
  • Sections that are in the backup replace the current data; sections that are not in it (for example spending, when restoring an old backup) are left alone.
  • A safety copy of your current data is written to backups/pre-restore-<timestamp>.json first. These files are never rotated away, so a mistaken restore can itself be undone by restoring the safety copy.
  • --dry-run shows the per-section before/after counts and writes nothing. Without --yes you are asked to confirm.
  • Data goes through the normal storage functions, so with LIFE_OS_ENCRYPTION_KEY set the restored files are encrypted. To restore an encrypted backup, set the same passphrase it was written with.

Health Check (Doctor)

hermes-life-os-doctor
hermes-life-os-doctor --profile alex --json
hermes-life-os-doctor --strict       # warnings also fail (exit status 1)

Hermes' storage layer is deliberately forgiving: a file it cannot read (wrong passphrase, corruption) quietly loads as empty so the app keeps running, which also hides the problem until the data is overwritten. The doctor is a read-only check that makes such problems visible and never changes anything. It looks at every data store (readable, valid JSON, right shape, consistent with the encryption key), the memory journal (unreadable lines, duplicate ids), leftover .tmp files from an interrupted write, backups (any at all, age of the newest, readable, complete, older ones unreadable), reminders (a time that can never fire, clashing ids, past one-offs) and budgets (bad limits, duplicate categories, out-of-range alert thresholds), the day-start setting (an invalid LIFE_OS_DAY_START_HOUR is flagged instead of silently ignored) entries dated in the future (a sign of a wrong computer clock or a bad import), persistent entry ids (rows that share an id, journal entries linked to a row that no longer exists) and the planner stores (a task due date, renewal date or birthday that cannot be read, an unknown body metric, an unusable savings target, a symptom without a valid severity, shared ids). Exit status is 0 unless there is an error (or, with --strict, a warning). The same report is available in chat as run_health_check.

Planner: Tasks, Body Metrics, Subscriptions, People & Today's Agenda

Until v1.42.0 Hermes tracked what you already did. Four new stores cover what is still ahead of you, and one tool pulls them together. All of it is local, per profile, encrypted with the rest of your data when encryption is on, and included in backup, restore, export and the doctor.

"add a task: renew passport, due friday, high priority"
"what's on my plate?"            "I did the passport thing"
"weight 80.5 today"              "how is my weight trending?"
"my blood pressure was 120/80"   "set a weight goal of 75 kg"
"add Netflix, 15 a month, renews on the 20th"
"I paid Netflix"                 "what do my subscriptions cost per month?"
"add Ayse, birthday 03-07, keep in touch every 14 days"
"I called Ayse"                  "who should I catch up with?"
"what's on my agenda today?"
  • Tasks (add_task, list_tasks, complete_task, update_task, delete_task, get_task_summary, clear_done_tasks): due dates (today, tomorrow, in 3 days, in 2 weeks, a weekday name or YYYY-MM-DD), priority, tags and repeats (daily, weekdays, weekly, monthly). Completing a repeating task moves it to its next due date instead of closing it. A task is found by id or by words from its text; an ambiguous match is reported rather than guessed.
  • Body metrics (log_body_metric, get_body_trend, set_body_goal): weight, body fat, waist, hips, resting heart rate, blood pressure (systolic and diastolic), blood glucose, temperature and SpO2, with unit conversion (lb to kg, F to C, mmol/L to mg/dL), plausible-range checks, a weekly rate of change from a least-squares fit, and goals with an estimated arrival date. Readings are backdatable, editable through edit_log and fed into the memory journal. This is a log of your own numbers, not medical advice.
  • Subscriptions (add_subscription, list_subscriptions, update_subscription, cancel_subscription, resume_subscription, delete_subscription, get_subscription_summary, pay_subscription): weekly, monthly, quarterly and yearly payments with the monthly equivalent of each, the total, and what renews soon. Dates that have passed roll forward and month ends are clamped (the 31st becomes the 28th). Paying a subscription also logs it as an expense in its category.
  • People (add_person, list_people, get_person, update_person, delete_person, log_contact, get_people_due): birthdays (MM-DD or YYYY-MM-DD, Feb 29 falls on Mar 1 in other years), an optional keep-in-touch goal in days, a contact history and who is overdue.
  • Today's agenda (get_today_agenda): overdue and due tasks, today's reminders, renewals within 3 days, birthdays within 7 days, overdue catch-ups, habit streaks that need a check-in today and budget alerts, in one answer. Empty sections are left out.

The proactive nudges now also mention an overdue task, a renewal within three days, a birthday and an overdue catch-up (at most two of these per run), and the doctor reports planner entries it cannot act on: an unreadable due date, renewal date, birthday or metric, and ids shared by two entries.

Lists, Savings Goals, Symptoms & Notes

Four more everyday trackers, all local, per profile, encrypted with the rest of your data when encryption is on, and covered by backup, restore, export and the doctor.

"add milk, eggs and bread to the shopping list"    "I got the milk"
"start a savings goal: holiday, 2400 by 2027-06-01"  "I put 300 into the holiday fund"
"headache, 6 out of 10, about 90 minutes, stress"    "how often did I have headaches this month?"
"note: gift idea for Ayse, a tea set #gifts"         "what notes do I have about gifts?"
  • Lists (add_list_items, get_list, get_lists, check_list_item, remove_list_items, clear_checked_items, delete_list): named checklists for things that are not tasks. The default is shopping; any other name (packing, movies, ...) works. Add several items in one go ("milk, eggs, bread"), tick them off while you shop, clear what is ticked afterwards. An item already open on the list is skipped instead of duplicated, and an item is found by id or by words from its text.
  • Savings goals (add_savings_goal, contribute_to_savings, get_savings, update_savings_goal, delete_savings_goal): a target, an optional deadline and starting balance. Add money or take some out (backdated if you like); you see the percentage, what is left, your real pace over the last 90 days (a starting balance does not count), the day that pace gets you there, and with a deadline whether you are on track and how much to put aside per month. A goal whose deadline is within two weeks shows up in the proactive nudges and in today's agenda.
  • Symptoms (log_symptom, get_symptom_summary): headaches, back pain, heartburn and the like with a severity from 1 to 10, an optional duration, suspected triggers and notes. The summary gives the count, average and worst severity, entries per week, days since the last one, the most common triggers, whether it is getting more or less frequent, the worst weekday and, when you log sleep, the average sleep on symptom days against other days. Symptoms join list_logs, edit_log, delete_log, undo_last_log, backdating and log_many. A personal log, not a diagnosis or medical advice.
  • Notes (add_note, list_notes, search_notes, update_note, delete_note): a quick-capture inbox with tags and pins; search finds notes containing all your words in the text or tags.

The doctor now also reports savings goals with an unusable target or deadline, symptom entries it cannot use, list items and notes without text, and ids shared by two entries; entries dated in the future are checked for body readings and symptoms as well.

Income, Debts, Countdowns & Supplies

Four more trackers, all local, per profile, encrypted with the rest of your data when encryption is on, and covered by backup, restore, export and the doctor.

"I got paid 3000 salary"                 "how did the last three months go?"
"I borrowed 500 from Ali, due the 15th"  "I paid Ali 200"
"Mehmet owes me 120 for dinner"          "who owes what?"
"count down to the holiday on 2026-12-20"  "how long since I quit smoking?"
"I have 30 vitamin D pills, one a day"   "when do I run out?"
  • Income and cash flow (log_income, get_cash_flow): income is a tracker like the others (list, edit, delete, undo, backdate, log_many). The cash flow report puts income next to your logged spending per month: net, the share of income you kept, the average of finished months and your income sources, and it says so when you have spent more than you earned this month.
  • Debts (add_debt, record_debt_payment, get_debts, update_debt, delete_debt): money you owe (owe) and money owed to you (owed, also "lent"). Record partial payments (backdated if needed; you cannot pay more than what is open), see both sides with a net figure, and overdue flags. A debt that is overdue or due within three days appears in the proactive nudges and in today's agenda.
  • Countdowns (add_countdown, get_countdowns, update_countdown, delete_countdown): days until an event, or days since one (past dates are fine). A yearly countdown always points at its next anniversary and counts the years; Feb 29 falls on Mar 1 in other years. Countdowns within two days are nudged, within a week they are in the agenda.
  • Supplies (add_supply, refill_supply, get_supplies, update_supply, delete_supply): pills, contact lenses, pet food, anything that runs out. Tell Hermes how many you have and use per day and it works out how many are left today and the day you run out. A refill adds to what is left (or sets the count); changing the daily use applies from today. Supplies running out within a week are nudged, within three days they are in the agenda.

The proactive nudges now show up to three planner items per run instead of two, and the doctor checks the four new stores (an income amount or date it cannot use, a debt without a usable amount or direction, a countdown or supply it cannot read, shared ids).

Local Days, Night Owls & Backdating

Every tracker stamps its entries with a calendar date, and everything that reads them back (streaks, "last N days", the budget month) has to agree on what "today" is. Before v1.39.0 the writers used your local clock while about a dozen readers used UTC, so for part of every day the two disagreed: a dose logged at 00:30 in Istanbul did not count towards today's streak, an expense logged just after midnight on the 1st landed in last month's budget, and in UTC+14 the streak was wrong most of the day. All of them now share one source, demo/clock.py, which is your local day.

The memory journal is deliberately different: its entries carry a UTC timestamp and the journal-based tools (life score, achievements, heatmap, correlations, "on this day") keep working in UTC.

Night owls: set LIFE_OS_DAY_START_HOUR (0-12, default 0) to the hour your day really starts. With LIFE_OS_DAY_START_HOUR=4, something logged at 01:30 counts for the previous calendar day.

export LIFE_OS_DAY_START_HOUR=4        # Windows: set LIFE_OS_DAY_START_HOUR=4

Backdating: forgot to log yesterday's lunch? The 13 log tools (log_meal, log_sleep, log_hydration, log_workout, log_stress, log_meditation, log_gratitude, log_focus_session, log_expense, log_social_interaction, log_substance, log_reading, log_medication) take an optional date (YYYY-MM-DD):

"I had a salad for lunch yesterday, about 350 calories"
"log 2 glasses of water for 2026-10-09"

Future dates, dates more than 365 days back and anything that is not a real date are rejected with a clear message and nothing is written. A backdated entry is also filed under that day in the memory journal. Backdated hydration only joins the log and leaves today's running counter alone.

Fixing Any Log: List, Edit, Delete, Undo

Expenses were the only thing that could be corrected until v1.40.0. A meal logged with the wrong calories or a workout logged twice stayed in every average, streak and correlation. Twelve more trackers now have the same tools (meals, sleep, workouts, stress, meditation, gratitude, focus, social, substances, reading, hydration, medication; v1.41.0 added hydration, whose running "today" counter is recomputed after a change):

"show my meals from the last 3 days"          -> list_logs
"that lunch was 450 calories, not 300"        -> edit_log
"delete the duplicate workout"                -> delete_log
"oops, undo that last log"                    -> undo_last_log

Every entry has a short id (shown by list_logs); edit_log takes the same field names as the matching log tool plus date, validates them and changes nothing if any is invalid. undo_last_log with no tracker removes whatever was logged last (medication has no journal entry, so name it explicitly); with a tracker name, or expenses, it removes that tracker's newest entry. Edits and deletes also fix the memory-journal entry the log wrote, and deleting a journal entry with delete_entry now removes the tracker row it was linked to. Older entries get an id the first time they are listed.

Catching Up on Forgotten Days

get_logging_gaps lists the days with nothing logged, so they can be filled in with the log tools' date argument. With no tracker it gives one line per tracker you actually use; with a tracker (including hydration and expenses) it lists every empty day. Only complete days count, today is not over yet, and it follows your LIFE_OS_DAY_START_HOUR. log_dream accepts date too.

"what did I forget to log this week?"         -> get_logging_gaps
"I slept 7 hours on the 9th"                  -> log_sleep with date 2026-10-09

Reviewing a Day, Searching, Duplicates and Batch Logging

Four more tools make the correction workflow practical:

"what did I log yesterday?"                   -> get_day_log
"when did I last have pizza?"                 -> search_logs
"did I log anything twice this week?"         -> find_duplicate_logs
"log breakfast, a 7h sleep and 3 glasses for the 10th"  -> log_many

get_day_log shows everything logged on a day, grouped by tracker with ids, ready for edit_log / delete_log. search_logs looks through food, notes, titles, names, triggers and categories. find_duplicate_logs reports identical entries on the same day (hydration is skipped, two identical glasses are normal); it never changes anything. log_many logs up to 50 entries in one call, each {tool, args} with an optional date: every shape and date is checked first, so one bad date means nothing is logged, and a failure inside one entry does not stop the others.

Habit Check-ins

update_habit is now one check-in per day. Checking the same habit in twice on a day no longer adds twice to the streak (and no longer writes a second journal entry), and a day missed in between restarts the streak instead of silently continuing it. A streak freeze still bridges a missed day. Like the log tools it takes an optional date: a check-in for another day is recorded in the journal, so the habit consistency analytics see it, but the running streak counters (which describe "up to today") are left alone.

Fixing Expenses

An expense used to be impossible to correct: a typo in the amount or a double entry stayed in the totals and the budgets for good. Every expense now has a short id (shown when you log it), and three tools manage them:

"show my expenses from the last week"           -> list_expenses
"that lunch was 14.50, not 41.50"               -> update_expense
"delete the duplicate coffee"                   -> delete_expense

update_expense can change the amount, category, note or date; invalid values change nothing. Updates and deletes also fix the matching memory journal entry, so the life score and "on this day" agree with the spending data. Expenses saved by older versions get an id the first time you list them.

Spending, Social & Substance Tracking

"spent 12 on lunch"
"hung out with my best friend for an hour, really good talk"
"had 2 cups of coffee this morning"

Three more trackers alongside nutrition/sleep/fitness/mental, added because a "life OS" that only tracks the body misses a lot of what actually moves the needle day to day:

  • Spending - logs expenses by category, and (like every other metric) feeds the correlation engine, so "do I spend more on stressed days?" is an answerable question, not a guess.
  • Social connection - time spent with other people and how connecting/fulfilling it felt (1-10) - loneliness and social wellbeing are as real a signal as sleep or stress, just rarely tracked anywhere.
  • Substances - caffeine, alcohol, or anything else worth watching, with amount and unit left free-form. Caffeine specifically feeds the correlation engine (e.g. against sleep quality); other substances are logged and summarized even if not yet wired into correlations.

Ask for a summary anytime: "how's my spending been this month?", "how much have I been socializing lately?", "how much caffeine have I had this week?".

Life Score & Achievements

"what's my life score today?"
"show me my achievements"

Life Score blends whatever you've logged that day - mood, sleep, hydration, stress (inverted), energy, focus - into a single 0-100 number with a plain-language label (Thriving / Doing well / Steady / Rough day / Tough day). Not a medical measure, just a transparent, at-a-glance way to answer "how am I doing overall" without mentally combining five numbers yourself - demo/life_score.py's components field always shows exactly which metrics fed the score, so it's never a black box, and a day with only one thing logged still scores fairly (missing metrics are excluded, not treated as zero).

Achievements (demo/achievements.py) are streak and milestone badges - a 7/30/100-day streak on any habit, a 7/30/100-day overall logging streak, and count-based badges ("first workout logged", "50 mood check-ins"). Entirely read-only and recomputed fresh every time - there's no separate achievements database to drift out of sync with your actual logs, so editing or deleting an entry updates progress immediately.

Wrapped - a shareable summary card

hermes-life-os-wrapped                              # last 30 days -> hermes-wrapped.png
hermes-life-os-wrapped --days 365 --out my-year.png --title "My 2026"
hermes-life-os-wrapped --days 7                      # a "your week" card

A single shareable PNG card - your average Life Score, entries logged, days active, average mood/sleep, your best day, and badges earned - in the spirit of Spotify Wrapped or GitHub's yearly contribution recap. Entirely local: reads only from data already on disk, makes no LLM or network calls, and the image never leaves your machine unless you choose to share it. Needs matplotlib (already a core dependency, same as the dashboard).

Reading & Medication Tracking

"read 25 pages of Atomic Habits for 20 minutes"
"took my vitamin D"
"skipped my omega-3 today"
  • Reading/learning (log_reading, get_reading_summary) - session count, total minutes, total pages over a window. Reading minutes also feed the correlation engine.
  • Medication/supplement adherence (log_medication, get_medication_adherence) - log a dose as taken or skipped, get an adherence percentage per medication over a recent window. Simple by design: no dosage/scheduling logic, no interaction warnings - just an honest log of what was actually taken.

Recommendations

"what should I focus on today?"
"any suggestions based on my data?"

demo/recommendations.py turns patterns already visible in your own data into concrete, actionable nudges - entirely local, rule-based, no LLM or network call involved, so every suggestion can be traced back to the exact numbers behind it:

  • Threshold nudges - e.g. average sleep under 6.5h or stress over 7/10 recently.
  • Correlation-derived insights - reuses the same correlation engine behind get_correlation_insights, just phrased as a suggestion.
  • Near-milestone streaks - "2 days from a 30-day streak on 'meditate' - keep it going!"

Not medical or therapeutic advice - a reflection of your own patterns, phrased as a nudge, nothing more.

What's New

v1.51.0 - Projects and Waiting For

  • New projects: add_project, get_projects, get_project, update_project, add_project_task, delete_project. A project is an outcome that needs several tasks; every project owns a tag and the tasks carrying it belong to it. Hermes shows progress, the next action and flags projects that are overdue or have no open task. The agenda lists project actions due today and a nudge points at stuck or overdue projects.
  • New waiting for list: add_waiting, get_waiting, chase_waiting, resolve_waiting, delete_waiting. Track what you need back from someone, with a follow-up date (a week out by default). Items due for a follow-up show in the agenda and nudges; chasing pushes the date out and resolving closes the item.
  • Recently deleted now labels waiting items by what you were waiting for.
  • Doctor, repair suggestions, search, backup, restore, export and recently deleted cover the two new stores.
  • 37 new tests - suite grew from 5017 to 5054.

v1.50.0 - Flashcards and Net Worth

  • New flashcards with spaced repetition (Leitner boxes): add_flashcards, review_flashcards, reveal_flashcard, answer_flashcard, get_flashcard_stats, delete_flashcard. A correct answer moves a card up a box (it returns after 1, 3, 7, 16 and 35 days), a wrong one sends it back to box 1 for tomorrow. The agenda and a nudge show how many cards are due.
  • New net worth: set_account_balance, get_net_worth, get_net_worth_trend, delete_account. Record what you own and owe whenever you look at it; Hermes keeps each account's history and shows net worth today, the change since the last update and a month-by-month trend. A nudge appears after five weeks without an update.
  • Recently deleted now labels flashcards by their front and weekly reviews by their week.
  • Doctor, repair suggestions, search, backup, restore, export and recently deleted cover the two new stores.
  • 39 new tests - suite grew from 4978 to 5017.

v1.49.0 - Inbox and Weekly Review

  • New inbox: add_to_inbox, get_inbox, process_inbox_item, delete_inbox_item. Capture anything the moment it crosses your mind, then turn it into a task, a note, a library item or one of today's priorities. The agenda shows what is unsorted and a nudge appears when items have waited three days or the inbox reaches ten.
  • New weekly review: start_weekly_review, save_weekly_review, get_weekly_reviews, delete_weekly_review. A guided end-of-week ritual: the facts of the week (tasks, priorities, time blocks, overdue, inbox, life wheel), three questions, a 1 to 5 rating and up to five focus items for next week that can become tasks. Friday to Sunday a nudge asks for it once you have done one.
  • The tool reference and the README tool count stay in sync through tests.
  • Doctor, repair suggestions, search, backup, restore, export and recently deleted cover the two new stores.
  • 38 new tests - suite grew from 4940 to 4978.

v1.48.0 - Daily Priorities, Plan My Day, Life Wheel, Library, Calendar Export

  • New daily priorities (the "Big 3"): add_priority, get_priorities, complete_priority, carry_over_priorities, delete_priority. At most three per day, free text or an open task (finishing the priority finishes the task), a streak of days with every priority done, and the agenda, the nudges and the closing summary show what is still open.
  • New plan my day: plan_my_day fits priorities, overdue and due tasks into the free gaps of your day plan. It shows a proposal first; apply=true saves time blocks linked to the tasks. Tasks take an optional minutes estimate (add_task, update_task).
  • New life wheel: log_life_wheel, get_life_wheel, delete_wheel_checkin. Rate eight areas from 1 to 10, see bars, the change since the last check-in and the lowest areas, and get a nudge after five weeks without a check-in.
  • New library: add_library_item, get_library, update_library_item, finish_library_item, suggest_library_item, get_library_stats, delete_library_item. One backlog for books, movies, shows, games, podcasts and courses.
  • New calendar export: export_calendar writes an .ics file (time blocks, task due dates, birthdays, countdowns, trips, expiry dates, chores, renewals, decision reviews) with stable ids, so a newer import updates earlier events instead of duplicating them.
  • New tool reference docs/TOOLS.md, generated from the schemas and checked by a test.
  • README restructured (quick tour, contents, demo); release history moved to CHANGELOG.md; added SECURITY.md and a pull request template.
  • Doctor, repair suggestions, search, backup, restore, export and recently deleted cover the three new stores.
  • 81 new tests - suite grew from 4859 to 4940.

v1.47.0 - Recently Deleted, Recommendation Ledger, Pinned Backups

  • New recently deleted: get_trash, restore_deleted, empty_trash. Every planner delete or clear tool now keeps what it removed for 30 days, so a wrong delete can be undone. A test fails when a new delete tool is not covered.
  • New recommendation ledger: get_recommendation_history, mark_recommendation. Suggestions from get_recommendations are filed; a followed one is compared with the metric before and after.
  • New pinned backups: pin_backup, unpin_backup. Pinned backups survive rotation and do not count towards the keep limit.
  • The doctor, get_repair_suggestions, search_everything, backup, restore, export and rekey cover the two new stores (trash and recs).
  • Ideas for these three came from a look at Matt Abate's life-distributable (approval gates, a recommendation ledger, backups that cannot be deleted by the agent); the code is our own.
  • 137 new tests - suite grew from 4722 to 4859.

v1.46.1 - Subscription renewals keep their day of the month

  • Fix: pay_subscription saved the clamped renewal date as the new schedule, so a monthly renewal on the 31st drifted to the 28th for good (Jan 31, Feb 28, Mar 28, Apr 28). Quarterly renewals from Nov 30 and yearly ones from Feb 29 drifted the same way. Every subscription now remembers the date its schedule was set from (anchor) and counts each renewal from it, so Jan 31 is followed by Feb 28, Mar 31 and Apr 30. Paying late, catching up after missed payments, resuming a cancelled subscription and the unpaid roll-forward all keep the day. Editing the date or the cycle by hand starts a new schedule from that date. A subscription that already drifted before this release stays where it is: set its real date once with update_subscription and it holds from then on.
  • Thanks to Matt Abate for the report (issue #34).
  • 24 new tests - suite grew from 4698 to 4722.

Older releases: see CHANGELOG.md.


Running Tests

python -m pytest tests/ -v
python -c "from environments.life_os_env import smoke_test; smoke_test()"

Why This Is Different

Every other agent in this hackathon does something for you. Hermes Life OS becomes something with you.

It tracks nutrition, sleep, fitness, stress, focus, hydration, habits, and goals - and connects them all. Bad Monday? It checks if you slept poorly Sunday. Energy crash at 3pm? It looks at what you ate for lunch. Mood dip this week? It finds the pattern you missed.

That is not a tool. That is a presence that accumulates.