uv tool install anamnesis-memory && anamnesis init

ἀνάμνησις (anamnesis): Greek for recollection, the act of calling knowledge back to mind.

Anamnesis is a local-first, file-based memory layer for Claude Code that syncs across your own machines. What Claude learns about your projects (conventions, architecture decisions, fixes that worked, what you did yesterday) is stored as plain markdown, indexed locally for search, and kept in sync over your private network.

Claude Code's own memory stays on the machine where it was written. Copying a SQLite index through a cloud folder such as Dropbox or iCloud corrupts it. Anamnesis syncs the markdown with git instead and rebuilds the index on each machine, so a note written on your desktop is searchable from your laptop.

No cloud account is required. The memory stays on your machines, version-controlled and human-readable.

How it works

  ┌─────────────┐        git over your private mesh        ┌─────────────┐
  │  desktop    │  ◄────────────  (Tailscale)  ────────►   │  laptop     │
  │             │                                          │             │
  │  Claude Code│                                          │  Claude Code│
  │     ▼       │                                          │     ▼       │
  │  MCP server │   markdown (source of truth)             │  MCP server │
  │     ▼       │   + SQLite FTS index (rebuilt locally)   │     ▼       │
  │ ~/.anamnesis│                                          │ ~/.anamnesis│
  └─────────────┘                                          └─────────────┘

• Files first. Each note is a markdown file under ~/.anamnesis/memory/, readable and git diff-able. The markdown is the source of truth; everything else is derived from it.
• A local index. A SQLite FTS5 index gives keyword (BM25) recall. It is rebuilt locally and is not synced. On the project's eval set, keyword search reaches about 94% recall, so there is no vector store until measurements call for one.
• Sync over git. The markdown is a git repo synced over your private Tailscale mesh, or any git remote you control. Only the markdown travels; each machine rebuilds its own index, so the database file is not copied between machines. If two machines edit the same note, sync reports a git conflict instead of dropping either side.
• Claude Code integration. An MCP server exposes read-only query tools, and session hooks do the rest: SessionStart injects the relevant notes, and SessionEnd captures a summary of the session and syncs it.
• Reflection and merge. An optional reflection pass (any OpenAI-compatible model, set in config) distills session notes into durable notes, and anamnesis merge consolidates near-duplicates. Generated notes carry provenance and confidence in their front-matter. merge --apply is gated on your eval set so it does not lower recall, and anamnesis eval measures reflection before and after on a sandbox copy.
• A dashboard. A web GUI to browse, search, edit, and read the history of your memory across every machine.

Quickstart

Prereqs: Claude Code, uv, and git.

uv tool install anamnesis-memory && anamnesis init

anamnesis init registers the MCP server with Claude Code at user scope, installs the SessionStart, SessionEnd, and PreCompact hooks, configures the store at ~/.anamnesis, and runs a first sync. It is idempotent: it backs up settings.json and does not duplicate hooks. --print shows the full plan without writing anything, and --local-only skips the remote until you want one.

Claude Code gets five tools: memory_search, memory_list, and memory_status (read-only, safe to auto-approve), plus memory_write and memory_sync. Full reference: CLI · MCP tools · configuration.

git clone https://github.com/oscardvs/anamnesis && cd anamnesis/server
uv venv --python 3.12
uv pip install -e ".[mcp,dev]"
uv run anamnesis init --print

Cross-machine sync

Memory is a git repo (~/.anamnesis/memory/) synced over your private Tailscale mesh, or any git remote you control. Set it up once:

1. Put every machine on the same tailnet (install Tailscale, tailscale up). Pick one always-on machine to host the shared repo; tailscale status prints its MagicDNS name (for example host.your-tailnet.ts.net).

2. Create one shared bare repo on the host:

   bash

   git init --bare -b main ~/anamnesis-memory.git
   

3. Point each machine at it:

   bash

   anamnesis init --remote 'you@host.your-tailnet.ts.net:anamnesis-memory.git'
   

   The host itself uses the local path: --remote "$HOME/anamnesis-memory.git".

Sync runs commit -> pull --rebase -> push and rebuilds the local index after pulling, so a note written on one machine is searchable on the others within a sync cycle. If you started with --local-only, re-run init --remote ... later; the store attaches to the remote and pushes its history.

Hooks

anamnesis init installs three Claude Code hooks:

• SessionStart injects the most relevant notes for the current project (your global preferences, the project's durable notes, and a couple of recent session summaries) and starts a background sync.
• SessionEnd captures an episodic note from the session transcript and syncs it, so it is on your other machines by the next session. PreCompact captures the same kind of note before context compaction.

Two more commands build on them:

• Reflection (optional). Point anamnesis config set reflection.provider ... at an OpenAI-compatible model and anamnesis reflect distills accumulated session notes into durable conventions. With reflection.auto set, it runs at session end once a project has enough unreflected notes. anamnesis merge consolidates near-duplicates and only applies if recall on your eval set holds.
• Import. anamnesis import copies Claude Code's own per-project memory into the store, so notes you already wrote for it are included.

To set the hooks up by hand instead of with init, copy examples/hooks.settings.json into ~/.claude/settings.json and point it at your install.

Benchmark

bench/cross-machine-tokens/ runs one scripted task on a fresh machine twice: once without Anamnesis, where the agent explores the project to learn its conventions, and once with the real SessionStart memory block injected. It drives the Claude Code CLI headlessly on a synthetic project, so a Pro/Max subscription is enough and no API key is needed.

The committed result.json (3 runs per arm on claude-opus-4-8) shows about 8% fewer total input tokens with memory, 366k versus 336k on average, with output tokens and turn count about the same. The conventions were known from the first turn instead of being rediscovered. An earlier version of the scenario used a 5-file project and showed no difference; that null result is documented in the same README.

Dashboard

A web GUI for the memory store: browse and full-text search every note, edit markdown with per-note history, see every machine (which one wrote what and when it last synced), and run reflection from the browser. Provenance badges show where each note came from: you, a session capture, reflection, or import.

npx anamnesis-dashboard      # http://localhost:3000

or, from the CLI you already have:

anamnesis dashboard

Needs Node 20 or newer. From a repo clone, cd dashboard && npm run dev works for development.

It is a thin read/write client over the same local store the MCP server uses: it reads the SQLite index directly and shells out to the anamnesis CLI for writes and sync. Use --port, --store, and --no-open to adjust how it serves. See dashboard/README.md for configuration and design notes.

Status

September 2026: v0.1.3 is on PyPI. The local-first core is done and has been used across a real multi-machine setup: the store, the MCP server, the hooks, git sync, the one-command install, the reflection and merge passes with their eval harness, and the dashboard. APIs may still change between releases.

Next: a hosted relay for people without their own mesh (the server already ships anamnesis relay commands behind a [relay] extra, but they are not documented yet) and team memory.

Repository layout

Contributing

Issues and discussion are welcome. If the install is rough on your machine, open an issue with the exact command you ran and its output.

License

Apache License 2.0. See NOTICE.