Claude Code's built-in memory
Claude Code has two memory systems, and both load at the start of every session. CLAUDE.md files are written by you: a managed policy file, a user file in ~/.claude/CLAUDE.md, a project file (./CLAUDE.md or ./.claude/CLAUDE.md) and a local, gitignored CLAUDE.local.md, plus path-scoped rules in .claude/rules/. Anthropic recommends keeping each file under 200 lines, and /init writes a starter one.
Auto memory is written by Claude. It is on by default, stores notes in ~/.claude/projects/<project>/memory/ with a MEMORY.md index, and loads the first 200 lines or 25 KB of that index each session, reading topic files on demand. It is machine-local, so it does not follow you to another computer or a cloud session. The /memory command lists every file, toggles auto memory and opens the folder. Source: Claude Code memory docs.
1. claude-mem — automatic session capture
What it is: captures what your agent does during sessions, compresses it with AI and injects relevant context into future sessions. Install with npx claude-mem install, or from the Claude Code plugin marketplace. It runs a local worker with SQLite and Chroma for vector search, and also documents Codex, OpenCode and Gemini CLI. Apache-2.0, about 96,000 GitHub stars on 4 October 2026.
Honest limits: many moving parts (a background worker, a vector database, hooks), and Windows users have reported a long run of issues on GitHub, for example failed installs looping, console windows flashing and vector search degrading. The installer also offers an optional hosted observer with a 14-day trial. Source: github.com/thedotmack/claude-mem.
2. Basic Memory — Markdown files you own
What it is: local-first memory stored as plain Markdown that forms a knowledge graph through observations and wikilinks. Connect it to Claude Code with claude mcp add, or use its plugin. It also works with Claude Desktop, Codex, Cursor and VS Code. The local version is free (AGPL-3.0); Basic Memory Cloud is $15 a month.
Honest limits: the local install needs uv and the --prerelease=allow flag; without it, the project says, you silently get an older release. Source: github.com/basicmachines-co/basic-memory.
3. Official Knowledge Graph Memory server — the minimal option
What it is: the reference memory server from the Model Context Protocol project: entities, relations and observations stored in a local JSONL file. One command: npx -y @modelcontextprotocol/server-memory (on Windows, launch through cmd /c). Works in any MCP client that runs local servers.
Honest limits: basic by its own description, with no semantic search described, and local only. Source: modelcontextprotocol/servers.
4. Mem0 — hosted memory with a Claude Code plugin
What it is: an open-source memory layer (Apache-2.0) with a hosted MCP server documented for Claude Code, Cursor, Codex, Windsurf and VS Code, plus a Claude Code plugin that captures memories through lifecycle hooks. Self-hosting is also possible.
Honest limits: needs a Mem0 Platform account; the free tier caps retrievals at 1,000 a month. The older OpenMemory app was removed in July 2026. Sources: Mem0 MCP, pricing.
5. Supermemory — hosted MCP with OAuth
What it is: a memory API with a hosted MCP server that signs in with OAuth and can both search and add memories; MIT-licensed, with self-hosting options. Free plan with $5 of credits a month.
Honest limits: credit-metered, and the older standalone MCP repository is deprecated, so follow the current docs. Sources: MCP docs, pricing.
6. mcp-memory-service — self-hosted, many backends
What it is: self-hosted persistent memory with a knowledge graph, a REST API and local embeddings, installed with pip and added to Claude Code with claude mcp add. Backends include SQLite, Cloudflare and Milvus. Apache-2.0.
Honest limits: to use it from claude.ai or ChatGPT you have to host it yourself on a public HTTPS endpoint with OAuth. Source: github.com/doobidoo/mcp-memory-service.
7. Graphiti — temporal knowledge graphs
What it is: Zep's open-source engine for real-time knowledge graphs, with an MCP server you run through Docker Compose. Apache-2.0.
Honest limits: the heaviest setup here: a graph database plus an LLM API key for ingestion. Source: getzep/graphiti MCP server.
8. BrainTube — a knowledge base Claude Code can search and write to
What it is: BrainTube is ours, so read this entry as labelled. It is a hosted MCP server (https://mcp.brain-tube.com/mcp, nothing to install) over the docs, videos, podcasts, articles and notes you save. Claude Code can search it and write to it, including bulk_ingest, which saves up to 50 items in one call. export_claude_md turns your library into a CLAUDE.md-compatible context file. The same endpoint works in Claude.ai, Cursor and other MCP clients. Free plan with MCP included (30 credits a month); paid from $9 a month.
Honest limits: it is a library of what you choose to save, not automatic session capture like claude-mem; usage is credit-metered; the server code is source-available under the Elastic License 2.0, not open source. Claude Code + bulk_ingest walkthrough · Connect guide.
Also worth knowing: Serena, a coding toolkit for Claude Code, includes a simple memory system as a side feature. Letta Code is a separate agent harness rather than a Claude Code add-on, so it is not listed.