macOS · Windows · ChatGPT and Claude on the web
Agent memory, shaped by the calls you make.
You decide things all day with Claude Code, Codex, ChatGPT and Claude: which draft to keep, which fix to turn down, what to check first. Agent Memory saves those sessions as you work and keeps each decision attached to the conversation it came from. Your next session, in any of those tools, can look it up.
Already using it? Sign in →
Why
What you decided yesterday is stuck in yesterday's session.
A new session starts with none of it. You explain the tone again, the naming rule again, the fix you already turned down. A notes file helps until it gets long, and it still loses the conversation that explained why.
Switching tools scatters it further. A correction you made in Claude Code never reaches Codex, and the newsletter tone you settled with ChatGPT is gone when you draft in Claude.
- MonChatGPTYou decided
No exclamation marks, and keep the newsletter under 200 words.
- TueClaudeBack to the old way
Big news!! Here's everything new this spring. (420 words)
- ThuChatGPTAsked again
Want me to add a few exclamation marks to make it livelier?
How it works
Saved as you work. Graded by your decisions. Ready for the next session.
Agent Memory runs next to the tools you already use, so there is nothing new to write down.
Every session, saved into one memory.
Claude Code, Codex, Devin CLI and Grok sessions are saved when they end, and a periodic sweep catches any that were missed. In ChatGPT and Claude on the web, you ask and the current chat is saved. The original conversation is kept, encrypted, so any answer leads back to where it came from.
SessionEnd hook · periodic sweep · Remote MCP
Graded by what you decided.
When you correct a draft, keep it or turn it down, your AI records that against the memory through MCP, and so does a look back weeks later. Search ranks with those records, in the cloud and on your computer, so the version you kept ranks higher.
feedback · decision review · preference pair
Start the next session from it.
Search what you already decided before you begin. Turn on the first-prompt hook and a new session in Claude Code, Codex or Devin CLI starts with what Agent Memory found. Save procedures, preferences and decisions by kind through MCP and pull them up again.
search · lifecycle hooks · first-prompt context · MCP
Illustration with sample data: sessions from five AI tools on one timeline. A decision made in Codex on September 12, marked adopted, is found and handed to a new Claude Code session.
Architecture
From every AI tool to your next session.
Where each part runs and where the data sits. Storage and embeddings are separate choices.
Features
What you get today.
Six AI tools, one memory
Claude Code, Codex, Devin CLI and Grok are captured by hooks. ChatGPT and Claude save over Remote MCP when you ask.
Provenance and replay
Every result points to its session. Read the conversation back in order, down to the message it came from.
Ranked by your decisions
Feedback, decision reviews and preference pairs move results up or down, in the cloud and on your computer.
Words and meaning
SQLite FTS5 and BGE-M3 vectors, merged with reciprocal rank fusion. Without vectors, word search still answers.
Context in the first prompt
With the UserPromptSubmit hook on, each new session starts with what Agent Memory found. Off until you enable it.
Embeddings on your device
BGE-M3 on MLX or onnxruntime. The model downloads on first use. Outside providers only with your own key.
Local mode
SQLite and sqlite-vec on this computer, with outbound traffic blocked. If the setting can't be read, it stays blocked.
CLI and MCP stdio
Thirteen tools for agents on the same machine, on Mac and Windows. No network port is opened.
Security
Security and where your data lives.
Only what we can check. This is how it works today.
Read how your data is handled →- Storage modes
- Cloud (default), local or hybrid, chosen in the app.
- Encryption
- Raw sessions are compressed and stored with authenticated encryption.
- Restore
- Back to the original JSONL, verified with SHA‑256.
- Per-user isolation
- Ownership is enforced at the PostgreSQL boundary. Nothing is shared between people.
- Local mode
- Outbound traffic is blocked, and stays blocked if the setting can't be read.
- Embeddings
- Made on your device: MLX on the Mac, onnxruntime on the CPU on Windows.
- Outside providers
- DeepInfra or Gemini only with your own key, never in local mode. There is no shared key.
- Error reports
- Never include conversation text, search terms, cookies, URL queries or screenshots.
- Thinking
- Kept for replay. Never searched, never used for memory.
- macOS
- Apple Silicon. Signed with a Developer ID and notarized by Apple.
- Windows
- Self-contained bundle, installed per version with rollback. Not code-signed yet.
Works with
Keep the AI tools you already use.
Captured automatically
- Claude Code
- Codex
- Devin CLI
- Grok
Remote MCP, when you ask
- ChatGPT
- Claude on the web
On your computer
- agent-memory CLI
- MCP stdio
Docs
Everything else is in the docs.
Install, sign in, pick a storage mode, connect ChatGPT and Claude, and turn on the hooks you want.
FAQ
Questions people ask first.
Who is it for?
Anyone whose work runs through more than one AI session: writing, research, analysis or code. It helps most when work moves from one session to the next, or from one tool to another.
Does it read my ChatGPT or Claude history?
No. On the web, Agent Memory saves the current conversation only when you ask, in a single call. It doesn't read your sidebar history or hidden reasoning.
Can I keep everything on my computer?
Yes. Choose local mode. Agent Memory blocks its own outbound traffic and stores everything in SQLite on the device. You can switch modes later in the app.
How does my feedback change results?
When your AI records that you adopted, rejected or corrected something, search takes it into account. Adopted memories rank higher, rejected ones lower, and a later review can reverse the verdict.
Do I need an API key for semantic search?
Not on your computer: BGE-M3 runs locally after a one-time model download. For semantic search through Remote MCP, add a query-embedding key of your own. Without one, search matches words.
Does it change how my agents behave?
Only through the hooks you turn on. Saving a session when it ends is on by default. Passing past context into the first prompt of a new session stays off until you enable it.
Can my team share one memory?
No. Agent Memory is personal. Memories are not shared between people.
Is it open source?
No. The source is private and the license has not been decided.
Get updates
Start your next session from what you already decided.
Agent Memory is opening up gradually, and there is no public download yet. Leave your email for product news and a note when you can start.
Already using it? Sign in →