Keep every Claude session attuned to how you like to work.
You've explained your style to Claude a hundred times — "be blunt," "skip the preamble," "I'm new to this part," "just show me the code." Compass remembers it for you: a small, curated profile of your working style that it safely splices into the memory every Claude session reads, so each one adapts to you from the first message.
- ~125 calibration questions, researched against Anthropic's steering docs + working-style psychology — each with best-first, click-to-pick answers.
- ~25 automatic guardrails — your answer becomes a rule Claude follows ("always confirm before deleting," "never commit secrets").
- You're in control — approve / edit / forget any note; a pause kill-switch that pulls Compass out of your sessions on the spot; nothing inferred goes live until you approve it.
- A light/dark dashboard showing the verbatim profile every session reads, a double-click app, and a system-tray companion.
- SessionStart auto-sync, built on the same audited safe-write engine as Lifejacket. 140+ tests.
Claude Code and Cowork start every session as a blank slate about you. Compass
fixes that. It keeps a profile of how you like to work and injects it into the
user-level ~/.claude/CLAUDE.md — read by both Claude Code and Cowork —
inside a clearly-marked, self-managed block. So "warm but concise," "challenge my
decisions," "I work ship-then-refine" travel with you into every new session, no
re-explaining.
It learns who you are two gentle ways:
- By asking. Now and then (never more than once every couple of days), Compass poses one thoughtful calibration question. Your answer becomes a profile note.
- From your history (v0.2, opt-in) — it will read your local Claude transcripts and draft profile notes for you to review.
The questions are researched, not guessed. The ~125 calibration questions are grounded in Anthropic's own prompt-engineering and steering guidance and in validated working-style psychology (Big Five, communication styles, learning & cognition, accessibility). Each offers best-first, click-to-pick answers, and ~25 set automatic guardrails — turning your answer into a hard rule Claude follows ("always confirm before deleting," "never commit secrets," "never use these libraries"). Small, high-signal, and chosen because each one measurably changes how Claude works for you.
Compass has a sibling, Claude Lifejacket. They're independent — each works perfectly on its own — but they're designed to complement each other, and together they're the magic:
| Lifejacket | Compass | |
|---|---|---|
| Syncs… | your projects | you |
| So every session knows… | what you're working on | how you like to work |
Install both and a fresh Claude session opens already knowing your projects
and your style — context and personalization, from the very first message,
with zero re-explaining. They write to the same CLAUDE.md but live in separate,
non-colliding blocks, so they never step on each other (or on your own notes).
Compass stores claims about you and will eventually infer some — so it's the most transparent and controllable tool you'll install:
- Nothing inferred reaches your sessions until you approve it. Things you say yourself go live immediately; anything Compass infers sits pending and is never injected until you review and approve it.
- See everything. The dashboard shows your profile grouped by category, where each note came from (you vs inferred), the verbatim text every session reads, and a full activity log.
- A real kill-switch.
compass pausedoesn't just stop syncing — it removes your profile fromCLAUDE.mdon the spot, so Compass instantly stops influencing every session, while keeping your data so you can fix andresume. - No ghost memories.
compass forgetremoves a note and re-syncs, so it's gone from your sessions too — not just the store.
It's built on the same carefully-tested safety engine as Lifejacket: it can't corrupt your memory file, it backs up before every change, it only edits its own labelled section, and it never overwrites notes you've hand-tweaked. 140+ tests back that up.
Easiest — the app (Windows): download Claude Compass.exe from the
latest release
and double-click it. A little window opens with your profile, the next question,
and Approve / Forget / Pause / Sync buttons. No Python, no terminal.
For the command line (any platform): Python 3.9+ (get it here — on Windows tick "Add to PATH").
- Windows: right-click
install.ps1→ Run with PowerShell, orpip install --user . - macOS/Linux:
pip install --user .
After installing, python -m claude_compass <command> always works; from the
project folder you can also run .\compass <command> (Windows) / ./compass <command>.
compass init # set up your local profile
compass quickstart # (optional) fill recommended defaults for a fast baseline
compass ask # answer a calibration question or two
compass answer comm_tone "warm but concise"
compass sync # share your profile with every Claude session
compass install-hook # make it automatic, forever
compass dashboard # see everything it knows, in your browser| Command | What it does |
|---|---|
compass init |
Create the local profile store (~/.claude-compass/) |
compass quickstart |
Fill the recommended answers for a strong baseline in one go |
compass ask |
Show the next calibration question |
compass answer <id> "..." |
Answer it (becomes a profile note) |
compass show |
Your full profile + where each note came from |
compass forget <n> |
Remove a note (and re-sync so it's gone everywhere) |
compass edit <n> "..." |
Edit a note in place (re-syncs) |
compass approve <n> / --all |
Approve an inferred (pending) note |
compass tray |
Run quietly in your system tray |
compass sync [--dry-run] |
Push your profile into Claude's memory |
compass pause / resume |
Kill-switch: pull / restore Compass's influence |
compass dashboard |
Open the visual status page |
compass status / log |
Where things stand / recent activity |
compass install-hook / uninstall-hook |
Turn automatic syncing on/off |
compass live on / off |
Live mode: re-inject your profile before every message (Claude Code), so edits land on your next prompt |
compass doctor |
Quick health check |
- v0.2 — learn-from-history: read your local Claude transcripts and draft profile notes (opt-in, you review before anything syncs); a system-tray companion; richer re-confirmation of stale notes.
Big AI products already have "memory" — so why Compass? Because the mainstream options infer things about you automatically and store them somewhere you can't fully see. Compass is built on the opposite principle. The fair comparison:
- ChatGPT Memory is the closest foil. It auto-saves facts and, since 2025, infers more from your history. It's wonderfully low-effort — but the common complaints (it "saves random things," the inferred layer is opaque, editing one thing doesn't reset the rest) are exactly what Compass sets out to avoid: nothing inferred reaches your sessions until you approve it, everything lives in a file you can read, and pause / forget take effect immediately. It's also ChatGPT-only; Compass feeds Claude Code and Cowork.
- Claude Memory —
Anthropic's own auto-synthesised summary — is great and built-in, but it's
claude.ai-only and the synthesis is opaque. Compass is a profile you curate,
in a file you own, that also travels into Claude Code via
CLAUDE.md. (If you'd rather not manage anything, Claude Memory is the zero-effort choice — that's a fair trade.) - mem0 / OpenMemory MCP is the power-user route: a universal memory the model queries across tools. It's auto-capture and pull-based; Compass is approval-gated and guaranteed-injected into the file every session already reads.
- Hand-written CLAUDE.md "voice/style" notes are the DIY version of this, and plenty of people do exactly that. Compass just automates and safeguards it: a managed block, a hook that refreshes it, and writes that can't corrupt the file or stomp your own notes.
If you want effortless, automatic memory and don't mind not seeing the wiring, the built-in options win. Compass is for people who want their working-style profile curated, visible, approval-gated, and instantly switch-off-able — feeding every Claude surface, not just one app.
- Claude Meter — live usage on your taskbar.
- Claude Lifeboat — backup & restore for your Claude data.
- Claude Lifejacket — keep every session aware of your projects.
- Claude Compass — keep every session attuned to you. (you are here)
- Claude Parachute — a safety net for the Bash changes Claude Code's
/rewindcan't see.
|
Built by Jack Bhanded, Lead developer and architect at SawYouAtSinai. Devotee of innovative technologies and gadgets. Built this because he was tired of re-explaining how he likes to work to every fresh Claude session. |
See CHANGELOG.md for the version-by-version list of changes.
MIT © Jack Bhanded — do whatever you want, just keep the copyright notice.