Docs
Overview
What Agent Memory keeps, and why it keeps decisions as well as conversations.
Agent Memory is a personal memory for AI agents. It keeps the sessions you run in Claude Code, Codex, Devin CLI and Grok, lets ChatGPT and Claude on the web search them, and hands what you decided earlier to the agent you are working with now.
Why it exists
Working with AI, you build up judgment: what to check first, which fix to reject, why a design was chosen. Almost none of it gets written down. Capturing it usually takes an interview, and the hardest part to keep is the context that gave each call its meaning.
Agent Memory captures that judgment as a by-product of the work. The original session is kept, so every decision stays attached to its context: what was asked, what the agent proposed, and what you corrected, adopted or rejected. The next agent you open starts from it, even when it is a different agent from the one where the decision was made.
How it works
- Capture. Sessions from Claude Code, Codex, Devin CLI and Grok are picked up when they end; ChatGPT and Claude save the current chat when you ask. Everything is normalized into one memory, and the original is kept for provenance and replay.
- Evaluate. What you corrected, adopted, rejected and later reviewed is recorded against the memory, and search ranks by it.
- Build on it. Search the memory as the next task starts, hand an agent the instructions and past decisions it needs through lifecycle hooks and first-prompt context, and save procedures, preferences and decisions by kind through MCP.
Concepts
| Term | Meaning |
|---|---|
| Session | One conversation with an agent, from start to end. Agent Memory stores the raw transcript and can replay it. |
| Message | One user or assistant turn. Search works on messages, and each result points to its session. |
| Memory | A distilled item with a kind: fact, preference, decision, procedure or episode. |
| Provenance | Where a result came from: the session, the message and the time. |
| Feedback | What happened to an agent's output: corrected, adopted or rejected, and why. See Decisions and feedback. |
| Workspace | The project a session belongs to. Search can be narrowed to one workspace. |
Where you use it
- Mac app: a menu bar app with a home view, sessions, replay, search and settings.
- Windows app: the same scope, including capture, storage modes, search and replay. See Windows.
- CLI and MCP stdio: for agents on the same computer. No network port is opened. See CLI and MCP stdio.
- Remote MCP: for ChatGPT and Claude on the web, at
https://agent-memory.jp/mcp. See Remote MCP. - Web console: sign in to browse your sessions and memories.
What it doesn't do
- It doesn't share memory between people. Agent Memory is your own memory.
- It doesn't read your ChatGPT or Claude chat history. On the web it saves the current conversation only when you ask.
- It doesn't search your agents' thinking. Recorded reasoning is kept for replay only.