# Documentation

- [Anamnesis documentation](/docs): A cross-machine, file-first memory layer for Claude Code. Markdown is the source of truth, synced over your own private network.
- Guide
  - [What is Anamnesis?](/docs/guide): A memory for Claude Code that persists between sessions and follows you across the machines you own.
  - [Install and connect to Claude Code](/docs/guide/install): Install Anamnesis, run anamnesis init, and start a new Claude Code session so memory turns on.
  - [What happens when you use it](/docs/guide/how-it-works): Memory is auto-injected at the start of a Claude Code session, saved as you work, captured when the session ends, and synced quietly in the background. You usually do nothing.
  - [Use it on all your computers](/docs/guide/across-machines): Connect your machines over Tailscale so your memory round-trips between Claude Code sessions.
  - [Browse your memory](/docs/guide/dashboard): Launch the dashboard to search, explore the memory map, read git-like history and diffs, view your machines, write and edit notes, and sync.
  - [Keep your memory tidy](/docs/guide/curating): Portable versus machine-local notes, editing and deleting in the dashboard, and how to reflect without losing work.
  - [FAQ and troubleshooting](/docs/guide/faq): Common gotchas, environment caveats, and how to fix them, in plain language.
- Internals
  - [Architecture](/docs/internals/architecture): The five layers (markdown source of truth, derived SQLite FTS5 index, git-over-Tailscale sync, FastMCP server plus lifecycle hooks, Next.js dashboard) and how Claude Code drives them.
  - [Notes, the markdown format, and the store](/docs/internals/data-model): The Memory record, the markdown-with-YAML-frontmatter format, the three note types, scope, the on-disk store layout, and the rebuildable SQLite FTS5 index.
  - [Keyword recall: FTS5 and BM25](/docs/internals/recall): How search() works: an FTS5 MATCH against memories_fts, BM25 ranking with a recency tie-break, OR-joined query tokens, and excluded superseded notes.
  - [Capture and injection (the hooks)](/docs/internals/capture-and-injection): How Anamnesis wires into Claude Code's lifecycle: what gets injected at SessionStart, and how a session becomes one episodic note at SessionEnd and PreCompact.
  - [Cross-machine sync over git and Tailscale](/docs/internals/sync): How the git sync backend works: the commit, fetch, rebase, push cycle, the conflict policy, what is and is not tracked, the Tailscale bare-repo topology, and why the SQLite index is never synced.
  - [Reflection and compression](/docs/internals/reflection): How Anamnesis distills episodic session notes into durable semantic and procedural memory: the swappable LLM, the reflect pipeline, redaction, the no-fallback policy, and the eval suite.
  - [The MCP server](/docs/internals/mcp-server): The FastMCP server named anamnesis: its five tools, their annotations, the pure functions behind them, and the concurrency model.
  - [Inside the dashboard](/docs/internals/dashboard-internals): How the Next.js memory GUI works. Read paths versus write paths, the API route surface, the memory-map data model, runtime requirements, and the app shells.
  - [Design decisions](/docs/internals/design-decisions): ADR-style record of the load-bearing choices: file-first over a knowledge graph, SQLite first-class with WAL, never sync the raw database, git-as-sync now, and a swappable reflection model.
- Reference
  - [CLI reference](/docs/reference/cli): Every anamnesis subcommand, its flags, defaults, and exact behavior, grounded in the source.
  - [MCP tools reference](/docs/reference/mcp-tools): The five Anamnesis MCP tools: exact names, parameters with defaults, return shapes, and read-only versus write (approval) status.
  - [Configuration and environment](/docs/reference/configuration): Every environment variable with defaults and purpose, the machine-local config.json fallback, the .mcp.json shape, and the filtered-environment caveat.
  - [Security and privacy](/docs/reference/security): What Anamnesis redacts before text reaches an LLM, what syncs versus what never leaves your machine, and the exact boundaries of those protections.