imece vs claude-mem
Comparison · Last updated: 10 October 2026
claude-mem is an open-source memory plugin for Claude Code: it captures what happens in sessions, summarizes it and adds it to later sessions. imece carries a team's decisions, their reasons and handoffs between agents. This page describes the difference as of 10 October 2026.
What each one is for
One helps your agent remember your sessions. The other helps the agents on a team remember what the team decided.
- claude-mem: tool usage in sessions is observed, summarized and added to the context of later sessions.
- imece: decisions, reasons, dropped approaches, unfinished work and work given to a role (handoffs) are recorded and reach your teammates' agents.
How claude-mem works
- Install:
npx claude-mem install, or inside Claude Code /plugin marketplace add thedotmack/claude-mem and /plugin install claude-mem. The README asks for Node.js 20 or later; Bun and uv are installed automatically if missing. Besides Claude Code, the README gives install commands for tools like OpenCode and T3 Code. - Hooks: the README lists five lifecycle hooks:
SessionStart, UserPromptSubmit, PostToolUse, Stop and SessionEnd. The architecture document gives the list slightly differently (for example, it has a PreToolUse hook for Read). - Capture: per the architecture document, the
PostToolUse hook captures tool executions and sends them to a worker for AI compression, the Stop hook generates the session summary, and the SessionStart hook adds context from earlier sessions. - Storage: SQLite (per the docs,
~/.claude-mem/claude-mem.db) and a Chroma vector database for search (optional in the architecture document). A local worker service provides an HTTP API and a web viewer. - Model: the README does not name the model used for compression. The project is built with the Claude Agent SDK, and you choose the provider: claude-mem's own observer (sign-in required, 30-day free trial), your own OpenRouter or Gemini key, or your Anthropic plan.
- Search and privacy: a
mem-search skill and MCP tools query the history. Content wrapped in <private> tags is not stored. - Multiple devices: Cloud Sync (cmem.ai Pro) replicates the local database across your own devices through a per-user sync hub. The documentation warns that if you turn it on, observation narratives and your full prompt text are uploaded to the sync hub under your account.
- Sharing: the Export/Import scripts in the docs export selected records as a JSON file and import them into another claude-mem installation. Sharing project-specific learnings across teams is among the listed use cases.
- License and version: Apache License 2.0. As of 10 October 2026 the latest release on GitHub is v13.35.0 (9 October 2026).
How imece works
- Writing: the agent records decisions, approaches it tried and dropped, unfinished work and work given to a role (handoffs), and it can search the team's records. Records are not generated automatically from tool usage; they exist when the agent calls a tool. Your teammates also write to the team channel and send handoffs from the desktop app.
- Reading: on Claude Code, imece adds the team's relevant records to the agent's context while it works.
- Handoff: it goes to a role. The agent of a member holding that role sees the handoff with its reason; they claim it, close it or reject it, and the outcome returns to the sender.
- Storage: records are kept on team members' devices and synced across the team. Team content is also kept on imece's servers in Germany. Records marked personal stay on your machine.
- Status: imece is in closed beta and free during the beta. The desktop app runs on macOS and Windows.
How imece works
Where they differ
- Whose memory: the claude-mem README describes memory as continuity across sessions, and the Cloud Sync documentation describes multiple devices tied to your own account. imece's records belong to a team from the start: there are members and roles, and records sync between members.
- What is recorded and who writes it: claude-mem captures tool usage observations through hooks, and an AI provider compresses them. In imece the agent writes the record itself through an MCP tool.
- Entering context: both add context at session start. claude-mem also offers a skill and MCP tools for search. imece adds the team's relevant records while the agent works, and the records are searchable.
- Sharing: in claude-mem, the documented way to share with someone else is an exported JSON file. In imece, records sync between team members and a handoff goes to a role.
- Where data lives: claude-mem keeps its local database under
~/.claude-mem, and cmem.ai sync is a separate option. imece keeps records on devices and also keeps team content on imece's servers; records marked personal stay on your machine. - License and price: claude-mem is Apache-2.0, and per the README offers a 30-day free trial of the claude-mem observer. imece is in closed beta and free during the beta.
Using them together
- claude-mem for the continuity of your own sessions, imece for the team's decisions and handoffs. They keep different records, and neither replaces the other.
- We have not tested running the two together on one machine. Both add to context through Claude Code hooks; watch how large your context grows.
imece setup claude-code keeps other tools' hooks in your settings files and only adds or removes its own entries (--uninstall removes them).
Limits
- The records imece adds are context: the agent sees them, and it can still go its own way. imece's hooks are not used to block a prompt or the end of a turn.
- The agent can skip recording. Writing depends on the agent calling a tool. imece's MCP instructions tell it to record in the turn a decision happens, but if it does not call the tool, no record exists. In claude-mem, capture is automatic through hooks.
- When a user asks about something unrelated, the agent may not bring up a pending handoff on its own.
- If the service is down or slow, nothing is added on that turn and the agent carries on.
- Records can go stale. imece flags only some of them; it does not catch every outdated record.
- You get into imece with an access request; you install claude-mem yourself with
npx claude-mem install.