MEMESH

Memory your AI coding agent doesn't lose between sessions

MeMesh is a memory layer that runs on your own machine, for Claude Code and other MCP-compatible tools. One SQLite file, no cloud, no API key by default — it captures architecture decisions, fixed bugs and lessons learned automatically, and surfaces what's relevant before the agent starts working. Open source, MIT licensed, free to use.

View source on GitHub →npm package page
95.60%
R@5 on a public, reproducible benchmark
One file
Local SQLite, no cloud
MIT
Open source, code fully public

SOUND FAMILIAR

Your AI agent doesn't just forget facts — it repeats work

  • It re-proposes the approach you rejected last month, in this exact conversation
  • It trips over the same failing test it hit before
  • It re-discovers the constraint that broke production in March
  • It asks you to re-explain the architecture it helped design

THREE THINGS

Not a long feature list, just what actually gets used every conversation

01

Captures automatically

Six hooks sit on Claude Code's own lifecycle: loading relevant memories when a session starts, checking for related lessons before a file edit, recording every commit, turning a session's fixed errors into a structured lesson when it stops, and rescuing knowledge before context gets compacted. Nothing to remember by hand — it accumulates from normal work.

  • Six auto-capture hooks across the whole session lifecycle
  • Failures become structured lessons automatically (error, root cause, fix, prevention)
  • Opt out any time: MEMESH_AUTO_CAPTURE=false

02

Finds the right memory

Keyword search (FTS5) is the default — no LLM, no API key — and it is good enough to find a memory about "OAuth PKCE" when you search "login security". Add an embedder (local Ollama or a cloud provider) and semantic search layers on top, ranked together.

  • FTS5 keyword search by default — zero config, zero cost
  • Optional semantic search via local Ollama or a cloud embedder
  • Handles CJK, Thai and other space-less scripts correctly

03

Works across tools

MeMesh is a standard MCP (Model Context Protocol) server, not a Claude Code-only plugin. The same memory connects to Claude Code, Codex CLI and Gemini CLI at once — a decision recorded in one tool is recallable from another. Hermes Agent (NousResearch) and OpenClaw integrate natively instead: each has a first-party plugin, at the same tier as their own built-in memory backends, with no manual tool calls needed — Hermes runs recall and remember automatically on every turn, while OpenClaw's auto-capture is threshold-gated (up to 3 memories per turn when triggered).

  • One SQLite file shared across multiple MCP clients
  • Verified working with Claude Code, Codex CLI, and Gemini CLI
  • A local HTTP API and CLI are open too, not locked to one interface
  • Native MemoryProvider plugin for Hermes Agent — automatic recall/remember every turn
  • Native memory plugin for OpenClaw — threshold-gated auto-capture, up to 3 memories per turn

GETTING STARTED

A few minutes to install, then mostly nothing to do by hand

  1. 01

    Install

    Run /plugin install inside Claude Code, or npm install -g @pcircle/memesh for a shell command.

  2. 02

    Connect to Claude Code

    The plugin wires itself in; the npm install needs one run of memesh install-hooks.

  3. 03

    Remember something once

    memesh remember "This project uses PKCE, not the implicit flow" — then memesh recall finds it.

  4. 04

    Let it run

    From here the hooks capture decisions and fixed errors on their own, and inject what's relevant at the start of the next session.

View source on GitHub →

No signup, no credit card. A local install is all it takes.

COST

Free and open source

Free, open source

MIT licensed, source fully public on GitHub. No paid tier, no usage limit.

Your data stays on your machine

Zero network calls by default: one local SQLite file, no cloud, no API key, nothing sent to any third-party server.

Advanced features are opt-in

Want an LLM to auto-generate lessons, tags or summaries? Bring your own API key. It is off by default, so calling it out is the only place a cost shows up.

YOU ARE PROBABLY WONDERING

Common questions

Is this a product PCIRCLE sells?

No. This is a tool we open-sourced and use ourselves every day building systems — free, no customization pricing, no support contract.

Do I need a cloud API to use it?

No. Keyword search is the default and needs no API key. Semantic search is opt-in and needs an embedder, local or cloud, that you configure yourself.

Does it only work with Claude Code?

No. MeMesh is a standard MCP server — Codex CLI and Gemini CLI are both verified to connect, sharing the same memory. Hermes Agent (NousResearch) and OpenClaw each have their own native memory plugin, sharing that same memory too.

Where does the memory live, and who can see it?

In one SQLite file on your own machine (~/.memesh/knowledge-graph.db). Nothing is sent to any server by default, including ours.

The source is on GitHub, and you can start today

npm install -g @pcircle/memesh, or /plugin install right inside Claude Code. Free, open source, and a local install takes minutes.

View source on GitHub →About PCIRCLE

MeMesh is built and open-sourced by PCIRCLE — the memory layer we use every day building AI Native systems.