I read Atomic Habits three times. Each time, same result: two weeks of tracking, then back to old patterns.
The problem isn't discipline. The problem is that manual tracking is itself a habit — and building a "habit of tracking habits" while building other habits stacks two hard things. The failure mode is architecturally designed in.
So I built the thing that was supposed to be too hard to automate.
Engram watches your work behavior automatically. You do nothing. It tells you who you actually are.
Not who you think you are. Who your data says you are.
This is real output from a real week. Not a demo. Not made up.
Focus Score: 6/10
329 commits across 12 repos this week.
High output — but chronic fragmentation.
You never fully left any codebase.
─── Pattern Detected ──────────────────────────────────────
Apr 13 produced 202 commits — 9.2x the weekly daily average.
The prior 4 days averaged 8 commits/day.
Something unlocked on Apr 12 that carried into Apr 13.
When you hit that mode, throughput multiplied by 9x.
───────────────────────────────────────────────────────────
Time Reality:
Planned deep work: ~6h
Actual deep work: 2 days (Apr 12–13)
Days at <10% capacity: 5 out of 7
Atomic Habits Lens (Make it Obvious):
You have a 9x day in your data.
You don't know what caused it. That's the problem.
Action: 2-line pre-work log this week. After 5 days,
correlate entries against commit counts. Find the trigger.
Open Loops: 8–10 threads started, not finished.
One Thing:
Figure out what was different on Apr 13.
Protect that condition. Everything else is secondary.
This is what happens when you stop trying to remember your patterns and start letting data show them to you.
Engram solves two problems simultaneously:
# Without Engram
You: Help me finish the auth refactor
Claude: Sure! What's the context? What have you done so far?
What files are involved? What approach are you taking?
# With Engram
You: Help me finish the auth refactor
Claude: You've been working on this since Tuesday. jwt.go and
middleware/session.go are the main files. You hit a
snag with token revocation — noted it as unresolved
in yesterday's session. Want to start there?
You didn't tell Claude any of that. Engram wrote it to ~/.claude/ last night.
Every habit tracker requires logging. Engram doesn't. It observes what you do and surfaces patterns you'd never find manually:
- The day that was 9x more productive than average (and why)
- The context switches that silently kill your output
- The tasks you start and consistently abandon
- The time windows where your work is actually high quality
Scenario 1: The morning handoff
You start a new Claude Code session on a complex feature. Instead of re-explaining everything, Claude already knows: what you built yesterday, what's still broken, what decision you made and why. Engram wrote a briefing at 23:45. Your AI picked it up at session start.
Scenario 2: The pattern you didn't know existed
After 30 days, Engram tells you: on days when you switch between 3+ projects, your commit output drops 60%. You thought you were good at parallel work. You weren't. You just believed you were. Now you block Tuesdays differently.
Scenario 3: The unfinished task that became a problem
A bug report comes in. You type @engram Have I seen this before? in Claude Code. Engram finds it: you hit this exact issue on March 15, traced it to deploy.yaml:47, left a note saying "check if this gets reverted." It did. Six weeks later. Now you know in 30 seconds instead of two hours.
Scenario 4: The weekly review that writes itself
Every Sunday you get a plain Markdown report. Focus score, patterns detected, open threads, Atomic Habits lens on your data. No logging required. Just a mirror of the week you actually lived.
You work normally
↓
Engram watches 7 data sources silently
↓
Every night at 23:45:
├── Collect: git, shell, apps, AI sessions, browser
├── Synthesize: Claude generates structured analysis
├── Store: plain Markdown in ~/your-memory-repo/
└── Bridge: key insights → ~/.claude/projects/memory/
↓
Next session: Claude Code reads the briefing automatically
Every Sunday: weekly behavioral report in your reports folder
Data sources:
| Source | What Engram collects |
|---|---|
| Git | Commits, repos, velocity, files changed |
| AI sessions | Claude / Cursor / Codex decisions and topics |
| Shell | Commands, tools used, patterns |
| Apps | Time per app (ActivityWatch or macOS native) |
| Browser | Research tabs, topics explored |
| System | Recent files, processes, projects active |
What it produces:
| File | What's inside |
|---|---|
consciousness.md |
Insights, mental model shifts, non-obvious patterns |
patterns.md |
Behavioral and work patterns, updated nightly |
weaknesses.md |
Recurring problems, anti-patterns |
tasks.md |
Open tasks extracted from sessions |
reports/week-*.md |
Weekly Atomic Habits-framed behavioral report |
James Clear's 4 laws — Engram automates all of them:
| Law | The manual version | What Engram does |
|---|---|---|
| Make it Obvious | Habit tracking sheet | Surfaces your real patterns automatically |
| Make it Attractive | Reward systems | Weekly report that's actually interesting to read |
| Make it Easy | Minimum viable habit | Zero logging. Zero friction. Runs while you sleep. |
| Make it Satisfying | Habit streaks | 30/60/90 day behavioral history, compounding clarity |
You can't change what you can't see.
Engram's thesis: the reason most people fail to implement Atomic Habits isn't lack of discipline — it's that observation requires effort, and effort requires a habit, and habits are what you're trying to build. It's turtles all the way down.
Remove the observation cost. Watch what happens.
See your own git data in 30 seconds. No install. No account.
curl -fsSL https://raw.githubusercontent.com/lessthanno/engram-agent/main/scripts/preview.sh | bashIt shows this week's commit velocity by day, your peak ratio, and which repos you touched. Your data. Instant.
curl -fsSL https://raw.githubusercontent.com/lessthanno/engram-agent/main/scripts/quickstart.sh | bash3 questions. 2 minutes. Never touch it again.
After a week, run this any morning:
$ python3 ~/engram-agent/mind_sync.py --report
╔══ engram · 2026-04-14 ══╗
▸ TODAY'S PRESCRIPTION
Yesterday: ✓
→ Protect the first 90 minutes. No meetings, no Slack until 10:30am.
Law: Make it Satisfying
Target: commits >=127
▸ OPEN TASKS
· FujiPay Stripe webhook (project: fujipay) [priority: H]
· ZKLM MVP scope definition [priority: M]
▸ WEEK 2026-W16
Focus: 4/10
Pattern: Apr 13 = 202 commits, 9.2x daily average. 5 zero days after.
One thing: Figure out what triggered Apr 13. Protect that condition.
▸ YOUR MODEL
+ Wednesday is your highest-output day (avg 42 commits)
- Days with meetings: 38% of your normal output
────────────────────────────────────────────────
@engram in Claude Code for deeper analysis
Manual install
git clone https://github.com/lessthanno/engram-agent.git ~/engram-agent
cd ~/engram-agent
bash scripts/install.shVerify: bash ~/engram-agent/scripts/verify.sh
Query your behavioral memory from any Claude Code session:
@engram What was I working on last Tuesday?
@engram Have I seen this bug before?
@engram What patterns am I showing this week?
@engram What are my open tasks?
@engram When am I most productive?
Installed automatically. Read-only. Never modifies your memory.
| Engram | mem0 / MemGPT | CLAUDE.md | Habit trackers | |
|---|---|---|---|---|
| Observes behavior | Yes | No | No | No |
| Zero manual input | Yes | No | No | No |
| Works if you forget | Yes | No | Yes | No |
| Finds unknown patterns | Yes | No | No | No |
| Bridges to AI tools | Yes | Yes | Yes | No |
| 100% local | Yes | Varies | Yes | Varies |
They're complementary. Engram is the behavioral data layer the others are missing.
- 100% local. No cloud sync. No external APIs required.
- Secrets scrubbed. Regex patterns remove keys and tokens before synthesis.
- You own everything. Plain text files. No lock-in. Delete anytime.
- Open source. Read every line. Trust nothing you can't verify.
What pattern did Engram find about you?
Real data only. No demos. The most surprising insight wins nothing except the respect of people who understand what 9.2x means.
Made with Engram? Add this to your README:
[](https://github.com/lessthanno/engram-agent)
- 8-source behavioral collector (git, claude, codex, cursor, apps, shell, browser, claude context)
- Claude synthesis with fallback chain (CLI → proxy → API → offline)
- Daily logs + 8 persistent analysis files
- Weekly behavioral report (Atomic Habits framing)
- Daily coaching prescription — tomorrow's one specific behavior change
- Personal behavioral model — output triggers + killers, data-backed, updates weekly
- Claude usage quality coach — context%, overflow detection, session depth scoring
-
@engramClaude Code agent (coaching, model, tasks, patterns, usage) - SessionStart hook injects today's prescription into every new session
- PreCompact / Stop hooks
- Interactive installer + verify + uninstall scripts
- Skill extension system
-
--reportflag — fast terminal coaching summary (no API, instant) -
--shareflag — shareable weekly behavioral card for Discussions / social - Zero-install preview script —
curl ... | bashshows your week in 30s
- Focus score trend graph (30/60/90 days, terminal sparkline)
- Calendar integration — scheduled vs actual deep work
- macOS menu bar app — prescription visible without opening terminal
- Linux support
- Cursor and Codex deeper integration
- Behavioral diff — "how did this week compare to your best week?"
- Team mode — aggregate patterns across a small team (opt-in, local)
- Behavioral diff — "how did this week compare to your best week?"
- Custom trigger definitions — "tell me if I'm doing X consistently"
- Export to Obsidian / Notion / Bear
Most productivity tools tell you what to do.
Engram tells you what you already do — and lets you decide what to change.
The difference: advice from a book is about someone else. Data from your own 30 days of behavior is about you. Generic patterns don't change people. Personal patterns do.
The long-term vision: a behavioral intelligence layer for knowledge workers that compounds over time. The longer you run it, the more it knows. After a year, it can tell you things about your work patterns that you couldn't figure out in a lifetime of journaling.
This is the external memory your future self will be grateful for.
- macOS (Linux/Windows roadmap Q3 2026)
- Python 3.10+ (pre-installed on Mac)
- Claude CLI (optional — adds AI synthesis; works offline without it)
- ActivityWatch (optional — adds app/keyboard tracking)
Zero pip dependencies. Zero npm. Zero Docker.
Plain Python, no frameworks, no build step.
Priority areas:
- Linux/Windows port
- More AI tool collectors (Windsurf, Zed, GitHub Copilot)
- Local dashboard (HTML report viewer)
- Calendar integration
Read CONTRIBUTING.md before opening a PR.
bash scripts/uninstall.shRemoves hooks and LaunchAgent. Your memory repo is preserved.
Engram is free and open source. Always.
The Pro tier is for people who want more coaching depth, longer history, and richer analysis.
| Free (OSS) | Pro ($10/month) | |
|---|---|---|
| Behavioral observation | Yes | Yes |
| Weekly Atomic Habits report | Yes | Yes |
| Daily logs | Yes | Yes |
@engram Claude Code agent |
Yes | Yes |
| Cross-month trend analysis | No | Yes |
| "Best day" pattern deep-dive | No | Yes |
| Prioritized feature requests | No | Yes |
| Private Slack for feedback | No | Yes |
Not sure yet? Run the free tier for 2 weeks first. If the weekly report surprises you, it's working.
Atomic Habits tells you to track your behavior.
Engram does it for you.