amem
amem is a local-only memory server that lets a coding agent store, retrieve, and manage durable project knowledge, reusable skills, and deferred tasks — all on your machine.
Recall context:
amem_contextreturns a small ranked memory packet for a query before you explore files or send big prompts.Write memory:
amem_rememberstores durable facts (constraints, gotchas, decisions) with kinds, file anchors, and workspaces;amem_supersederetires outdated claims.Learn the contract:
amem_recipeexplains when to read vs. write memory.Inspect state:
amem_reposlists monitored repos/workspaces,amem_graphdumps stored facts/components/flows,amem_statusshows binding, policy, and counts, andamem_statsreports usage and token savings.Manage skills:
amem_skill_listshows a cheap index,amem_skill_viewloads a full procedure, andamem_skill_savestores a new multi-step skill.Track tasks:
amem_task_list,amem_task_add,amem_task_update, andamem_task_completemanage a Backlog → Next → Doing → Blocked → Done Kanban board.Measure value:
amem_usage_reportrecords which anchors you actually opened so savings become a real measurement.
Registers the local amem MCP endpoint for GitHub Copilot and generates a marked block in .github/copilot-instructions.md, instructing the agent to query the personal memory graph instead of re-discovering decisions and file ownership each session.
Binds the repo to amem's local MCP server for JetBrains IDEs (amem init --platform jetbrains), exposing memory search and task-board tools so IDE agents start oriented with durable facts about the codebase.
Provides guidance for Neovim-based agents through the shared AGENTS.md instruction file, telling the agent to consult amem's local memory and task board before broad codebase exploration.
Adds amem's generated instruction block to AGENTS.md for Warp agents, so they query the local memory graph (claims, anchors, savings tracking) rather than re-grepping the tree each session.
amem
Personal agent memory that stays on your machine. Works with Claude Code, Cursor, Claude Desktop, Copilot, Codex, Gemini CLI, Windsurf, Continue, Aider and Zed, on macOS, Linux and Windows.
npx @iamem/amem setup --platform claude # or cursor, codex, copilot …
Website · Features · Why amem · How it works Coding agents forget between sessions. They re-read the same files, re-learn the same constraints and spend tokens rediscovering decisions you already made.
amem gives them a private memory that lives on your machine. The next session starts with what the last one learned.
What you get
Memory: short facts (constraints, gotchas, decisions, owners) tied to the files they describe. Facts are flagged stale when those files change.
Context on autopilot: hooks hand your agent a small, ranked packet at session start and with each prompt. No pasting a project brief.
Skills: repeatable procedures (release steps, deploy order, test setup), loaded only when they apply.
Tasks: a Backlog → Next → Doing → Done board your agents update, so follow-ups survive the chat.
A dashboard:
amem uishows what your agents remember, what needs review and open tasks. Retire or delete a fact in one click.Private by design: one SQLite file in
~/.amem, a localhost-only UI, optional encryption and local backups.
Works with Claude Code, Cursor, Claude Desktop, Copilot, Codex, Gemini CLI, Windsurf, Continue, Aider, Zed and any other MCP client.
Related MCP server: io.github.oleksiijko/pmb-ai
Quick start
cd ~/code/your-project
npx @iamem/amem setup --platform claude # or: cursor, codex, copilot, windsurf … (amem platforms)
amem ui # open the dashboard at http://127.0.0.1:7843Setup installs the amem CLI, adds the hooks, registers the amem MCP server (restart your client once) and seeds the repo with build, test and CI facts from its manifests, so the first question already has an answer. From then on your agent checks memory before exploring and saves what it learns. You can also do it by hand:
amem remember "Stripe webhooks must be verified with the raw body" --kind gotcha --anchor src/webhooks/stripe.ts
amem context "stripe webhook signature"### claim.remember_2f41c0c9e1ff
Kind: `gotcha`
Why: `keyword+12`, `fts+4.2`, `embed+5.4`, `kind:gotcha`, `fresh`
Stripe webhooks must be verified with the raw body.
Anchors: `src/webhooks/stripe.ts`Any MCP client can use the same memory: amem mcp --print-config prints a ready-to-paste config.
Something off? amem doctor --fix rewires the current repo's client. To remove amem, run amem uninstall to see what it added, then amem uninstall --yes. Your memory stays unless you add --purge.
How well it finds things
npm run bench loads 36 memories from a sample web shop and asks 32 questions phrased the way people ask, often without the memory's words. It also sends 8 vague prompts ("ok sounds good, go ahead") that name no topic. CI fails if these numbers drop.
Path | Right memory first | Right memory in top 3 | Vague prompts given memory |
Prompt hook (automatic) | 88% | 97% | 0% |
| 81% | 94% | n/a |
The data is in bench/retrieval.json. Add your own cases and run npm run bench -- --misses to see what it gets wrong.
Privacy (non-negotiable)
Memory never leaves your machine: no sync, no cloud embeddings, no "share with org" mode.
The UI and HTTP MCP endpoint bind to
127.0.0.1only, and~/.amemis created with mode0700.Credentials and tokens are blocked from being saved. Personal notes stay out of work projects unless you ask about them.
Optional AES-256-GCM lock (
amem lock) and encrypted local backups (amem backup schedule).No network calls by default. If you opt in (
amem setupasks once, oramem telemetry on), amem sends an anonymous ping at install and first run, then at most one a day with the version, OS and which parts ran (MCP, hooks, UI, CLI). No memory, repos, paths or names.amem telemetry offstops it; managed (IT policy) installs never send it.
Learn more
Full guide: setup, the day-to-day loop, what gets stored, skills, savings, other clients, every command
License
MIT, see LICENSE.
Available Tools
16 toolsamem_contextA
Retrieve compact local amem memory for a query. Use before exploring files or calling a remote LLM with a large prompt. Pass workspace for named app memory, not only git repos.
Inputs: query (required), platform, sessionId, workspace.
amem_graphA
Dump stored facts for a workspace or git repo: claims, components, flows, and recent activity. Use when asked what amem remembers about a project. Prefer amem_context for a ranked packet for the current question. All arguments are optional.
Inputs: days, workspace.
amem_recipeA
Return the generic amem remember contract: when to call amem_context vs amem_remember. Use if you are unsure whether to read or write memory. Same recipe for every MCP host. Takes no arguments.
Inputs: no named parameters.
amem_rememberA
Store a durable local fact in amem (stays on this machine). Use kind private (or anchor private) for personal, non-work facts so they never appear in work repositories' context. When the fact replaces an older one (a newer status, a fixed bug, a changed decision), pass the old claim ids in supersedes so the outdated memory stops being served. The response lists possiblySupersedes: older look-alikes to retire with amem_supersede if they are now wrong.
Inputs: text (required), kind, anchors, workspace, +1 more.
amem_reposA
List what amem is monitoring. Each item has kind git or workspace. Workspaces also have a stable slug (MCP id); the display name can differ. Use when asked which projects, workspaces, or repos amem knows about. All arguments are optional.
Inputs: kind.
amem_skill_listA
List stored skills (procedures the agent learned) as an index of names and descriptions only. Cheap — call this first, then amem_skill_view to load the one you need. Skills are longer procedures; use amem_context for small durable facts. All arguments are optional.
Inputs: query.
amem_skill_saveA
Save a non-trivial multi-step procedure you just worked out as a reusable skill. Use after solving something worth repeating — a workflow, a dead end you found the way past, or a correction the user gave you. Provide a full SKILL.md in content, or a description plus body.
Inputs: name (required), content, sessionId, description.
amem_skill_viewA
Load the full body of one skill by name, or a supporting file inside it. Call this only after amem_skill_list (or the 'Relevant skills' section of an amem context packet) shows a skill worth following.
Inputs: name (required), file, sessionId.
amem_statsA
Usage and savings stats (local lookup time, estimated tokens/ms saved, hit rate) for one workspace/repo or everything. Use when asked how much amem helped, how often it hit, or usage over time. All arguments are optional.
Inputs: days, scope, workspace.
amem_statusA
Binding, policy, and counts for one workspace/repo, plus the full monitored inventory. Omit workspace for a machine-wide overview. All arguments are optional.
Inputs: workspace.
amem_supersedeA
Retire outdated claims so they stop appearing in context: marks them superseded (kept for history, not served). Pass by with the claim id that replaces them, if any.
Inputs: ids (required), by, workspace.
amem_task_addA
Create a deferred agent task on the project Kanban (default backlog). Use for work that should not get lost in chat. Durable facts still go through amem_remember.
Inputs: title (required), body, status, anchors, +1 more.
amem_task_completeA
Mark a Kanban task done as soon as its work is finished and verified, so the next session does not pick it up again. Records the completion time and keeps the task in history (amem_task_list with include_done shows it). Same as amem_task_update with status done. Returns the completed task as JSON; an unknown id returns an error.
Inputs: id (required), workspace.
amem_task_listA
List deferred agent tasks (Kanban) for a workspace/repo. Use for work to do later — not durable facts (use amem_remember for those). Default excludes done tasks. All arguments are optional.
Inputs: status, workspace, include_done.
amem_task_updateA
Change an existing Kanban task: move it between columns (set status to doing when you start it, blocked when it waits on someone), rename it, or replace its notes or file anchors. Only the fields you pass change. Use amem_task_complete to finish a task and amem_task_add for new work. Returns the updated task as JSON; an unknown id returns an error.
Inputs: id (required), body, title, status, +2 more.
amem_usage_reportA
Report what you actually did with the last amem context packet, so its savings estimate becomes a measurement instead of a guess. Call this once after you have finished acting on a packet. Do NOT estimate tokens — just list which of the anchors amem gave you you ended up opening anyway (anchors_opened), and whether the packet answered the question without further exploration (answered). amem computes the saving from the real size of the files you did not have to open. Reporting that you opened everything is a useful and expected answer.
Inputs: answered (required), anchors_opened (required), event_id, workspace.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 16 tools
Tools are mostly distinct, but amem_context, amem_graph, amem_status, and amem_stats all touch the 'what does amem know' space, and amem_task_complete is explicitly documented as amem_task_update with status=done. Descriptions do the heavy lifting to draw boundaries (e.g. amem_recipe as a routing decision aid), so an attentive agent can pick correctly, but the overlaps require reading.
Every tool is prefixed with a consistent amem_ namespace, grouped sensibly (skill_*, task_*). The pattern mixes bare-noun names (amem_repos, amem_stats, amem_graph, amem_status) with verb_noun names (amem_task_add, amem_skill_save, amem_remember), so it deviates slightly from a single convention but remains readable and predictable within subclusters.
16 tools is at the upper edge, but the server spans four coherent subdomains (memory, skills, tasks, stats) so each tool carries weight. There is some redundancy (amem_stats vs amem_usage_report, amem_task_complete vs amem_task_update), but nothing gratuitous.
Core lifecycle is well covered: memory (remember/context/graph/supersede), skills (save/list/view), tasks (add/list/update/complete), plus reporting. Minor gaps exist—no delete/rename for skills, no delete for tasks, and no explicit fact-delete beyond supersession—but these are partly mitigated by the completion/supersede semantics.
Maintenance
Related MCP Connectors
Shared memory for coding agents. Stop re-explaining your codebase every session.
Private memory for AI agents: save and recall facts across Claude Code, Codex, Cursor and ChatGPT.
Private, portable memory and reusable skills for AI agents.
Persistent memory for AI agents, for code, research and everyday life. Stale facts retire.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceGives AI coding agents persistent memory by storing observations, decisions, and learnings in a local SQLite database with vector search, full-text search, and a rules engine.4MIT
- AlicenseAqualityBmaintenanceLocal-first memory for AI coding agents using SQLite, enabling persistent recall of decisions, lessons, and project context without cloud or API keys.10139 PyPI289Apache 2.0
- FlicenseNot gradedqualityAmaintenanceProvides a local-first persistent memory layer that lets AI coding agents store, recall, and semantically search project-specific and cross-project knowledge across sessions, so context, decisions, and guidelines survive between conversations. Runs entirely on the user's machine with SQLite and no API keys, and optionally supports autonomous task-queue execution built on the same memory.-
- FlicenseNot gradedqualityAmaintenanceProvides a local-first, cross-session memory layer that lets AI coding agents store and semantically search per-project decisions, preferences, error patterns, and research alongside global guidelines. Also supports an optional task queue for autonomous execution, all backed by on-disk SQLite with no API keys or external calls.-