mcp-memory: SQLite FTS5 + Google's OKF for Fast Agent Memory

Looking for a way to give your AI coding agent persistent memory without bolting on a heavy vector database? mcp-memory is an open-source MCP server that stores agent memory as Google's OKF v0.2 Markdown files, indexes everything locally with SQLite FTS5, and exposes four clean MCP tools. It's designed to drop into Claude Desktop, Cursor, Antigravity, Windsurf, or Codex with a zero-boilerplate setup wizard.

What it does

mcp-memory maintains a dual-layer architecture:

• Human-browseable OKF directory: Every memory is dumped as a raw .md file inside memory/, with a root index.md (versioned with okf_version: "0.2") and a log.md for update history.
• SQLite FTS5 index: Full-text search and triggers keep key lookups under 20ms and keyword searches instant.

Memories are stored as OKF v0.2 Markdown documents with YAML frontmatter (type, key, namespace, tags, generated, sources, verified, status, stale_after).

MCP tools

The server exposes four primary tools:

• memory_store – Stores or updates a record. Requires key, content, and project_root. Optional: tags, namespace (default default), concept_type, title, description, resource, status, stale_after, sources, verified, generated_by.
• memory_retrieve – Gets a specific memory by key and namespace.
• memory_search – Finds memories by query, tags, or namespace. Supports limit (default 10).
• memory_get_last – Retrieves the last session checkpoint (system/last_memory) so the agent knows where work left off.

Namespace isolation and setup

Memories are scoped by namespace (e.g., user/preferences, project/architecture), keeping contexts separate. Setup is a single command:

python3 setup.py

This auto-configures installed MCP tools (Antigravity, Claude, Cursor, Windsurf, Codex). The server also supports a project_root parameter for every tool, so you can keep per-project memory stores.

Who it's for

Developers using AI coding agents (Claude, Cursor, etc.) who want persistent, searchable memory that stays in plain-text Markdown and doesn't require a vector DB or external service.