🌐 English | 繁體中文

MeMesh

Shared local memory for Claude Code and Codex.

MeMesh helps your coding agents remember your preferences, project decisions, and lessons from past work, so you spend less time explaining the same things in every new session.

• Remember across sessions. Save a decision or lesson and find it again later.
• Share context between tools. Claude Code and Codex on the same computer can use the same memories.
• Leave a handoff. Keep a note of where work stopped for the next session to check.
• Exchange messages. Running local agents can send each other work, questions, and results.

Memories live in one file on your computer: ~/.memesh/knowledge-graph.db. MeMesh needs no cloud memory service or separate AI model. Automatic reminders depend on the host integration; a handoff is context, not a guarantee that a task will restart itself.

Install

Claude Code — run these inside Claude Code, then restart it:

/plugin marketplace add PCIRCLE-AI/memesh
/plugin install memesh@pcircle-memesh

Codex — run these in your terminal, then restart Codex:

codex plugin marketplace add PCIRCLE-AI/memesh
codex plugin add memesh@pcircle-memesh

To use the command-line tools too, install Node.js 22.13+, then run:

npm install -g @pcircle/memesh
memesh doctor

Installation and update guide · Other integrations

Use it

Ask your agent in plain language:

Remember that this project uses SQLite.

What did we decide about login last time?

Save where we left off so the next session can pick it up.

Claude Code's plugin provides automatic capture and reminders. The Codex plugin can provide context at session start once its hooks are allowed to run; other capture behavior varies by host. You can also ask the agent to store or look up memories directly.

Full tool and command reference · Agent messaging · How it works · Contributing

Tools and integration details

All 12 Memory and Coordination Tools

Tool	What it does
remember	Save a memory; decisions need a reason.
recall	Find memories.
forget	Archive a memory or remove one observation.
learn	Save a mistake, its cause, and how to avoid it.
briefing	Read the project's current memory context.
task_state	Read or record the goal, next step, blocker, and done work.
message	Find running local agents and exchange messages.
export	Back up memories as JSON.
import	Import a memory backup.
work_package	Prepare a bounded memory digest or transcript package for human review.
user_patterns	Summarize work patterns.
improvement	Propose an improvement for human review.

---

Plugins ship 10 hooks. The intent reminder recognizes "remember this" (5 languages). The included Dashboard has 4 tabs, 11 languages.

Scored Ranking — relevance (30%) + recency (25%) + frequency (18%) + confidence (17%) + recall impact (10%). See the architecture for details.

The fine print

Scored Ranking — Results ranked by relevance (30%) + recency (25%) + frequency (18%) + confidence (17%) + recall impact (10%).

Agent messaging, the exact rules (full guide: docs/platforms/agent-messaging.md):

• Works today: an MCP, HTTP, or CLI sender can durably send one untrusted JSON-encoded payload of at most 65,536 UTF-8 bytes (64 KiB) to one named local recipient. A receiver can fetch it separately, resume from an opaque cursor after restart, and record intake, acknowledgement, workflow disposition, and host activation as separate facts.
• With the MeMesh Codex plugin enabled, each startup or resumed ordinary Codex CLI thread with a valid thread identity and existing working directory registers automatically under a thread-scoped identity; no manual agent setup is required. SessionStart launches an owner-private companion. SessionEnd keeps a bounded 45-second idle queue window, resume replaces the prior exact generation, and expiry removes the registration. A message accepted during that idle window reaches the thread as a short notice when the same thread resumes, and the agent then fetches the body from the inbox; it is not a claim that a stopped UI was awakened. memesh agent setup codex-session remains available only when one workspace needs a stable named principal. The complete native envelope, including routing metadata and payload, is capped separately at 16,384 bytes (16 KiB). An exact-session send returns success only after that native queue accepts it; an oversized full envelope reports native_message_too_large, an unreachable local router reports router_unreachable, and other unavailable or rejected sessions report recipient_unavailable. Scoped recovery data remains durable for all sender-side and recipient-side failures. Principal targets retain durable store-and-forward behavior. Native acceptance is not acknowledgement or disposition, and native messages must contain no secrets.
• A stopped, missing, or disconnected Codex session is not woken up or replaced, and a failed exact-session native delivery is not replayed automatically; the sender must retry deliberately. Its scoped recovery data stays available; memesh message storage report shows what is stored. Native delivery works on macOS and Linux only.
• This documented native path covers ordinary Codex CLI. Do not assume Codex Desktop or an unattached task registers unless that exact running session appears in message discover; this is an evidence boundary, not a claim that those hosts are universally incompatible.
• Pairing Claude Channel with automatic Codex registration needs no --project value copied between the two: each host derives its own routing project from its own working directory (Claude) or --workspace (Codex), so starting both in the same repository lands them in the same project automatically.

MIT license