Projects by folder
Memories belong to the folder you work in, or are Global. One repo's rules never reach another, and Claude Code worktrees count as their repo.
AutoMem: One memory for all your AI coding tools
AutoMem learns your preferences, project conventions and decisions from your conversations, and keeps them in a library on your computer that you can read and edit. Then it adds the relevant ones to later requests, in Claude Code or Codex, without changing either tool.
Claude CodeYou, in my-app
From now on, use redb, not SQLite, in this project.
AutoMem learns it
Library Saved
From now on, use redb, not SQLite, in this project.
my-appLearned
Added to the next request
CodexLater, in my-app
Add a cache for parsed files.
+ 1 memory from AutoMem
The problem
Each tool keeps its own notes
Claude Code and Codex each keep their own notes, and neither knows what you told the other. Switch tools, or start a new session, and you explain it all again.
Or you copy instruction files such as CLAUDE.md and AGENTS.md between tools by hand.
What you end up repeating
- Preferences“Keep answers short. Show the code first.”
- Project conventions“This repo uses redb, not SQLite.”
- Decisions“We chose WebSockets over polling. Don't switch back.”
How it works
A small proxy on your own computer
AutoMem sits between your AI tools and their providers, only for the hostnames you turn on. Your requests go straight to your provider; there is no AutoMem server in between.
Turn on your tools
Choose Claude Code, Codex or both in Settings, then select Turn on. AutoMem points only their provider hostnames to itself.
Work as usual
Your tools keep their normal URLs and your own accounts. Each request goes on to the real provider, and the reply comes back unchanged.
AutoMem learns
Clear statements in your messages become memories, saved under the project folder or as Global.
Both tools remember
Later requests, in either tool, get the relevant memories as one marked note on your latest message.
Features
Memory that stays out of your way
A library you can read and edit
Add, edit, search or delete any memory. Each learned memory shows the message it came from.
Learning your way
Built-in rules save clear statements like “Always…” or “From now on…”. Or try the experimental local model (Qwen3 0.6B, on your CPU), or your own model endpoint.
Prompt caching keeps working
Each message keeps the memories it got when it was sent. Earlier messages never change, so the tools' prompt caches still work.
Side requests left alone
Subagents, titles, summaries and safety reviews pass through untouched. They get no memories, and nothing is learned from them.
Your tools stay as they are
No plugin, no MCP server, no new URL. Claude Code and Codex keep their settings and your own accounts.
Windows, macOS and Linux
A portable Windows app you unzip and run, plus builds for macOS and Linux.
One native Rust binary
No Electron, browser or JavaScript runtime. One small executable, with pure-Rust TLS.
Screenshots
The app
Security and privacy
What AutoMem does on your computer
To add memories, AutoMem reads your AI tools' HTTPS traffic on your own computer. Here is what that means, in plain words.
A new certificate authority on each computer
AutoMem makes its own certificate authority when you turn it on. The private key stays in memory and is never saved, so it never leaves your computer. Turn off removes the certificate again.
With Start with Windows (Start at login on macOS and Linux), the key is kept so your tools keep trusting it. It is protected by Windows (DPAPI), the macOS keychain or a Linux keyring; without a keyring, only by file permissions.
Only the hostnames you choose
AutoMem handles only the provider hostnames of the tools you turn on: api.anthropic.com for Claude Code, and api.openai.com and chatgpt.com for Codex. Nothing else goes through it.
Other apps that use those hostnames pass through too, unchanged. The ChatGPT website and app share chatgpt.com with Codex.
Turn off undoes everything
Turn off restores your hosts file and removes the certificate AutoMem added. If AutoMem stops unexpectedly, a separate helper undoes the setup, and automem --recover cleans up after a crash or power loss.
With Start with Windows on, the setup stays while AutoMem is closed, so those tools can't connect until it starts again or you run --recover.
Your data stays on your computer
Memories and captured conversations stay in a library on your computer, as plain text (not encrypted). Login headers and cookies are never saved. Common secrets such as API keys are removed from captured text, but not every sensitive passage can be recognised.
There is no AutoMem account and no AutoMem server.
What leaves your computer
Your AI provider gets your request as usual, with the added memories, over HTTPS that AutoMem checks against the provider's real certificate. To find the provider, AutoMem asks Cloudflare's public DNS (1.1.1.1).
Built-in learning and the local model run on your computer. If you set your own learning endpoint, it gets selected text from your messages and existing memories of the same scope, with its own key, never your tools' credentials.
You can check all of it
Activity shows every request: its project, the memories added and the exact context sent. Each learned memory shows the message it came from.
You can edit or delete any memory, or export the whole library.
Download
Get AutoMem
Free, with no account. Unzip and run: there is no installer.
Windows
x64- File
AutoMem-0.10.1-windows-x64.zip- SHA-256
771f167b5f3bcdbb658aafc1760ca50d56cfff439af34f96691323e9618576e8
- Unzip the whole folder, then run
automem.exe. - SmartScreen may say “Windows protected your PC”, because the app has no publisher signature yet. Choose More info, then Run anyway.
- Needs 64-bit Windows 10 or 11 and a CPU with AVX2 (most PCs from 2015 on).
macOS
Apple Silicon- File
AutoMem-0.10.1-macos-arm64.zip
- For Apple Silicon Macs (M1 and later). Unzip, then open
AutoMem.app. - The first time, macOS blocks it because it isn't notarized yet. Then open System Settings → Privacy & Security and choose Open Anyway. Or run this in Terminal:
xattr -dr com.apple.quarantine AutoMem.app - Some tools, such as Node.js ones, ignore the system's certificates. Start them with
NODE_EXTRA_CA_CERTSset totls/ca.pemin AutoMem's data folder (Settings shows where).
Linux
x64- File
AutoMem-0.10.1-linux-x64.zip- SHA-256
43c361f32e14d4587b9d66a0b5bc12cbc85bbbc17971ca6f1a7d8fc0e9e257fc
- Unzip, then run
./automem(if needed, firstchmod +x automem). - Needs an X11 or Wayland desktop. Turn on asks for your password through pkexec (polkit).
- Some tools, such as Node.js ones, ignore the system's certificates. Start them with
NODE_EXTRA_CA_CERTSset totls/ca.pemin AutoMem's data folder (Settings shows where).
Try it safely
- Open it and look around first. Your system settings don't change until you select Turn on, and every client stays off until you turn it on.
- Turn on asks for administrator permission, to trust AutoMem's certificate and point the chosen hostnames to it.
- Restart your AI tools if they were already running, so they trust the new certificate.
- Turn off undoes everything. After a crash, run
automem --recover.
Check the download
The result should match the SHA-256 on the card.
- Windows (Command Prompt)
certutil -hashfile <file> SHA256- macOS
shasum -a 256 <file>- Linux
sha256sum <file>
Status
Where it stands
AutoMem is new. This is what works today, and what comes next.
Works today
- Claude Code and Codex on Windows, in daily use by its developer since late October 2026.
- Learning from clear statements, with an optional local model or your own model endpoint.
- Recall by project and Global scope, without breaking prompt caching.
- Turn off undoes every change, and a recovery command cleans up after a crash.
- 668 automated tests, passing on Windows and Linux.
- Linux: builds and runs. Tested in Debian and Ubuntu containers (the window, system setup and recovery), not yet on a desktop.
- macOS: builds, but not yet tested on a Mac.
Next
- Learning from everyday work. The built-in rules catch only clear statements, and most coding chats have few. Next: decisions and corrections inside tasks, and the instruction files you already keep.
- Sync across your devices, as an optional paid service. Each computer keeps its own certificate; only memories sync.
- Shared project memory for teams, so a repo's conventions and decisions reach every teammate's tools.
- Testing on real Macs and Linux desktops.
Updated 30 October 2026.