MEMESH

Shared memory and clear handoffs for your AI agents

MeMesh 4.8.0 is a local collaboration layer for AI agents. Shared memory, durable exact-recipient messages and improvement proposals live in one SQLite file. Claude Code, Codex, Cursor, Gemini and custom or locally hosted agents can use it through MCP, HTTP or CLI — vendor neutral, open source, and no cloud required.

View source on GitHub →npm package page
4.8.0
Durable local agent collaboration shipped
One file
Memory, messages and receipts in local SQLite
Vendor neutral
MCP, HTTP and CLI — no single AI vendor required

WHEN YOU USE MORE THAN ONE AGENT

The problem is no longer just memory — it is the handoff

  • A decision Claude Code remembers has to be explained again in Codex
  • A review request stays trapped in the conversation where it was written
  • A system says a message was delivered without showing whether it was fetched, accepted or completed
  • A useful agent finding gets buried in a log because nobody decides what happens next

FOUR THINGS

From remembering work to coordinating it

01

Memory follows the work

Architecture decisions, fixed bugs and lessons stay in a local knowledge base. A new session or a different agent can recall what matters before continuing.

  • Six Claude Code hooks capture memory during normal work
  • FTS5 keyword search works by default with no API key
  • Optional semantic search can use local Ollama or a cloud embedder

02

Messages have recipients and receipts

Version 4.8.0 adds durable, exact-recipient local messages. A temporarily offline recipient does not make the message disappear. Send, fetch, acknowledge and workflow disposition remain separate, auditable facts.

  • Send through MCP, HTTP or CLI
  • Messages stay in local SQLite with cursor recovery
  • Queue admission is never presented as read, accepted or complete

03

Vendor neutral by design

The core is not a proprietary format for one AI vendor. Claude Code, Codex, Cursor, Gemini, custom agents and locally hosted models can use the same memory and messaging surfaces through open interfaces.

  • MCP for compatible agent hosts
  • HTTP and CLI for custom workflows
  • Native wakeup depth varies by host while the core local data stays shared

04

Humans govern improvement

Agents can turn memories and evidence into improvement proposals, but they cannot approve their own ideas. Only a human can accept or reject a proposal, keeping learning actionable without silently turning it into product policy.

  • Proposals retain links to source memories and evidence
  • Acceptance, rejection and current status stay visible
  • An accepted proposal remains unverified work until separately implemented and checked

GETTING STARTED

Start with shared memory, then add collaboration when you need it

  1. 01

    Install

    Install the MeMesh plugin in Claude Code, or run npm install -g @pcircle/[email protected] for the command-line tools.

  2. 02

    Connect your tools

    Use MCP with Claude Code, Codex, Cursor or Gemini. Custom workflows can use HTTP or CLI.

  3. 03

    Share one memory

    Store one project decision and recall it from another agent to confirm they use the same local memory.

  4. 04

    Send a handoff

    When work needs an owner, use message with an exact recipient; MeMesh keeps delivery and workflow receipts separate.

View source on GitHub →

No signup and 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 and no usage limit.

Your data stays on your machine

Memory, messages and receipts stay in local SQLite. No cloud or API key is required by default.

You control storage

Inspect message usage, preview pruning before applying it, and set an optional quota. MeMesh does not silently delete unresolved messages.

YOU ARE PROBABLY WONDERING

Common questions

Is this a product Partners Circle 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.

Is MeMesh tied to one AI vendor?

No. MeMesh exposes shared memory and collaboration through MCP, HTTP and CLI. Claude Code, Codex, Cursor, Gemini and custom or locally hosted agent workflows can use the same local data. Native wakeup depth varies by host, but the core is vendor neutral.

Does sending a message mean the recipient read it?

No. MeMesh records send, fetch, acknowledgement and workflow disposition separately. Queue admission never stands in for read, accepted or complete.

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

Memory, messages, receipts and proposals live in one SQLite file on your own machine (~/.memesh/knowledge-graph.db). Nothing is sent to our servers by default.

One local collaboration layer across your AI agents

Install @pcircle/[email protected], share memory across the agents you already use, and add durable exact-recipient handoffs when the work needs an owner.

View source on GitHub →About Partners Circle

MeMesh is built and open-sourced by Partners Circle — the local memory and collaboration layer we use while building AI Native systems.