Memory
Durable facts (constraints, gotchas, owners, how-tos) ranked by FTS, on-device embeddings, pins, and freshness. Stale anchors get marked. You approve drafts — amem does not silently invent truth.
What amem actually does
Coding agents forget between sessions. They re-grep the same tree, re-learn the same constraints, and burn tokens rediscovering decisions you already paid for. amem is a private local sidecar so the next session starts oriented.
The problems
| Without it | With it |
|---|---|
| Agent explores broadly every session | It queries memory first, then verifies the right files |
| Decisions die in chat scrollback | They become claims with file anchors and a Why line |
| The same deploy dance is rediscovered | A skill loads only when that work comes up again |
| “Do this later” vanishes when the tab closes | A Kanban the agent can add, move, and complete |
| Pressure to dump context into a shared wiki | Explicitly personal — prompts and learnings stay on the laptop |
| A flat AGENTS.md that goes stale | A small graph: components → flows → claims, updated as drafts |
What you get
Skills
Multi-step procedures as SKILL.md files. Agents see names and descriptions only, then load a body with amem_skill_view. Session-end can suggest one worth writing. Content is scanned before it is stored.
Tasks
Backlog → Next → Doing → Blocked → Done. Open tasks ride in the context packet. Use memory for facts, the board for “do later.” Works across every repo you track.
Local UI + desktop
amem ui is Setup, Dashboard, Memory, Analytics, Tasks, Skills — localhost only. amem app is the same thing in a window. Optional login item so it comes back after reboot.
Hooks and MCP
Cursor and Claude get a project rule plus stop hooks. Any MCP host can call amem_context, amem_remember, skill and task tools. Same SQLite for Windsurf, Continue, Aider, Zed.
Lock, backup, policy
Optional AES-256-GCM lock, encrypted local snapshots, a daily timer. Enterprise policy.toml can deny export, lock platforms, and stage skill writes. amem doctor --attest is the IT ticket packet.
How a day goes
- You (or the hook) run a context lookup. Matching claims, open tasks, and skill names land in the prompt.
- The agent prefers those file anchors over a broad explore, then verifies the current code — memory can be stale.
- When something durable is learned, it remembers a fact or saves a skill. Follow-ups go on the board.
- At session end, amem may queue a compact memory draft, a miss→learn draft, or one skill suggestion. You approve in the UI.
Privacy, on purpose
- The tool is public (GitHub / npm). Your database is
~/.amem/graph.dband is not shared. - The Cursor rule you commit is guidance only — no memory contents.
- The UI binds to
127.0.0.1. There is no “share with org” mode and no cloud embed API. - Optional anonymous install ping (package version, Node, OS). Opt out with
AMEM_TELEMETRY_DISABLED=1.
Not this
- Not a company wiki or a hosted RAG product.
- Not an agent runtime — Cursor and Claude still do the work. amem is the librarian.
- Not a replacement for reading the code. Trust fresh claims; re-check anything marked stale.
One command. Then the UI.
npx @iamem/amem setup
npm i -g @iamem/amem && amem setup
Node 20+. Then run amem ui. Source on GitHub.