Your LLM's memory lives in Markdown, not a black box

What it does

Basic Memory is an MCP server that acts as a shared notebook between you and your AI tools. It stores project context, observations, and wikilinks as ordinary Markdown files on your disk, readable by both humans and agents like Claude, Cursor, or Codex. When a conversation restarts, the AI searches its accumulated notes instead of asking you to re-explain the codebase.

The interesting bit

Rather than locking memory inside a proprietary vector store, it keeps everything as plain text that Obsidian can open and edit directly. Every exposed MCP tool carries behavior hints—read-only, destructive, idempotent—so agents choose correctly without burning context on trial and error. Semantic search runs locally with FastEmbed embeddings over SQLite, so the whole stack works offline.

Key highlights

• Local-first AGPL project; optional $15/mo cloud tier with cross-device sync and mobile access
• Native MCP integration with Claude, Cursor, Codex, VS Code, ChatGPT, and anything else that speaks the protocol
• Bidirectional editing: humans write in Obsidian while AI agents read and append the same files
• Hybrid full-text and vector semantic search backed by SQLite or Postgres
• Consolidated source tree also ships plugins for Claude Code, Hermes, and OpenClaw

Caveats

• Local-only users must handle their own cross-device sync (Git, Syncthing, etc.) and get no mobile app
• Some third-party integrations (Hermes, OpenClaw) currently rely on temporary install paths while upstream host support lands
• The repository bundles multiple agent packages alongside the core server, increasing the surface area to audit

Verdict

Grab it if you routinely waste the first ten minutes of a Claude session re-explaining your project. Skip it if you don’t use an MCP-compatible client or already have a note-taking system that your AI can read and write.