Skip to content
ai-memory
EnglishPortuguês (Brasil)Españolעברית日本語한국어
Menu

Product

Solutions

Integrations

Developers

Language

Get startedGitHub

Integrations

Your agents, your OS, your providers

27 coding agents and editors, four operating system targets, and a list of LLM and embedding providers that you can ignore, because everything works without one.

Find your agentGo to quick setup

Many agents, one memory

Each agent connects to the same server. What Claude Code learned this morning is there when you open Codex this afternoon.

Hub and spoke diagram: Claude Code, Codex, Cursor, Gemini CLI, OpenCode and Zed all connect to one memory in the centre.

Agents and editors

Filter by status, then pick a row to get its install commands.

Claude Code Supported

Claude Code
# launch it through ai-memory; the first launch installs its hooks and MCP entry
ai-memory run claude

# or wire it by hand, for a harness you never launch this way
ai-memory install-mcp --client claude-code --apply
ai-memory install-hooks --agent claude-code --apply

Hooks and MCP. Optional session-aware routing. Leave out --apply to print the change without writing it.

Supported coding agents and editors
AgentStatusNotesai-memory run
SupportedHooks and MCP. Optional session-aware routing.Managed
SupportedHooks and MCP. Session end is wired from Codex CLI 0.145.0.Managed
SupportedMCP config and lifecycle hooks.Direct launch
SupportedMCP config and lifecycle hooks.Direct launch
SupportedV1 or V2, detected from the executable.Managed
SupportedForces V2 when you do not want detection. Every finished turn refreshes the summary and the handoff.Managed
SupportedCapture works. The handoff arrives with the first tool call.Managed
SupportedTen hook events and MCP. The handoff arrives in Copilot’s additional context at session start.Direct launch
SupportedHooks and MCP, including post-compaction capture.Direct launch
SupportedTen hook events. Handoffs inject on the first prompt.Managed
SupportedBoth the v2 and v3 engines. Closes on exit under ai-memory run.Managed
SupportedMCP and four hook events, with native resume. Closes on exit under ai-memory run.Managed
SupportedUnder ai-memory run the session closes on exit. Otherwise, finalize-session.Managed
SupportedA generated extension gives capture and an MCP bridge.Managed
SupportedNative MCP and a TypeScript extension.Managed
SupportedMCP config and native plugin hooks.Direct launch
SupportedCapture includes subagent events. In Zero’s global sandbox settings, allow network and writes to the ai-memory data directory, or deliver queued events with ai-memory hook-drain from a normal shell.Direct launch
SupportedSix triggers. Handoff injection works.Direct launch
SupportedA native macOS agent app with its own ai-memory bridge. Nothing to install.Direct launch
Managed-onlyLaunch it with ai-memory run crush.Managed
Hooks-onlyPaste a generated hooks snippet into its settings.Direct launch
MCP-onlyChat uses mcp-remote on macOS, Windows and Linux (beta). Local Code-tab sessions use Claude Code hooks; install them with ai-memory install-hooks --agent claude-code --apply. Cowork capture is unverified.Direct launch
MCP-onlyCopilot exposes no lifecycle hooks yet.Direct launch
MCP-onlyRegistered as a context server.Direct launch
MCP-onlyProject-level MCP config.Direct launch
MCP-onlyUser-level MCP settings.Direct launch
SupportedThe hook installer prints a block for ~/.hermes/config.yaml. Recall and handoffs come from the community plugin.Direct launch

What each status means

Supported
Automatic capture through lifecycle hooks, plus the memory tools over MCP.
MCP-only
The agent can search and write memory. It exposes no lifecycle hooks, so nothing is captured automatically.
Hooks-only
Sessions are captured automatically. The agent has no MCP client to query memory with.
Managed-only
Works when launched through ai-memory run. There is no separate hook installer.
Community
A plugin maintained by the community, not installed by ai-memory itself.
Experimental
It ships and it works, with rough edges you should expect.

Per-agent details, config file locations and known quirks are in the support matrix and the MCP install guide.

Hooks capture, MCP answers

These are two separate connections, and a fully supported agent has both. The status labels above say which ones your agent gets.

Diagram with two lanes. Hooks: your agent streams events into memory on its own, which is automatic capture. MCP: your agent sends a request to memory and gets an answer back, which are tools the agent calls.

Hooks

The agent’s lifecycle events fire a small command that records prompts, tool calls and session boundaries. You do nothing, and the agent spends no tokens on it. Secrets are stripped before anything is stored.

MCP

The agent gets memory tools it can call when the conversation needs them: search, read a page, write a page, accept a handoff. MCP-only agents can use memory fully, but nothing is captured unless they write it.

Let Claude Code run alone, inside a sandbox

Even with --dangerously-skip-permissions, Claude Code keeps asking, so someone sits and babysits it. Switching every prompt off is a real risk. ai-memory run --yolo names what it is about to disarm, and when ai-jail is usable it offers to re-run the session inside it. --true-yolo adds Claude Code’s bypass mode on top; your own ask rules and its command-safety checks still apply.

  • The warning says what changes

    No prompt means the agent can delete files, run any command and reach the network. You read that before you say yes.

  • ai-jail boxes the session in

    bubblewrap with Landlock and seccomp on Linux, sandbox-exec on macOS. Say yes and a checklist picks which host credentials the jail gets. Say no and the session still runs, outside the box.

  • Only a person is asked

    The questions appear on a real terminal. Hooks, CI, detached and piped runs are never interrupted, and a session already inside ai-jail is not asked twice.

How the launcher worksAbout ai-jail
your terminal
ai-memory run --true-yolo claudeINFO ai_memory_cli: ai-memory starting version="2.6.5" bind=127.0.0.1:49374--yolo runs every tool call with no confirmation. An agent can delete files, run any command, and reach the network unsupervised.Proceed? [Y/n] yai-jail is installed. Re-run this session inside it? [Y/n] y

Linux or macOS with ai-jail installed. --jail goes straight in with sensible defaults and needs no --yolo; --no-jail keeps the warning and skips the offer.

Operating systems

One Rust binary per platform. On Windows, WSL2 is the supported path and the native build is experimental.

Supported operating systems
PlatformStatusHow it installsNotes
LinuxSupportedRelease tarball, Docker or Podman image, an AUR package or a Fedora RPMx86_64 and ARM64. Native binary, Docker or Podman, AUR packages with systemd units.
macOSSupportedRelease tarball with a launchd agent, Docker, or the menu bar app built from sourceApple Silicon and Intel. Native binary with a launchd agent, Docker, or a menu bar app you build from source.
Windows via WSL2SupportedThe Linux binary, inside WSL2Follow the Linux path inside WSL2 and launch your agent from there.
Native WindowsExperimentalRelease zip, Docker Desktop wrapper, or a source buildRelease zip with ai-memory.exe, the Docker Desktop wrapper, or a source build. The zip upgrades itself with ai-memory upgrade.
NixOSSupportedFlake package, or the NixOS module as a system serviceFlake package and a NixOS module with a hardened systemd unit. Secrets through agenix, sops-nix or an environment file.

Download the latest releaseRead the Windows guide

LLM and embedding providers

Both are optional. With neither, hooks still capture, search still works, and every session still ends with a summary page and a handoff.

Embeddings: local by default

Since 2.0, vector search runs in-process with a small open model. It needs no API key and no GPU. The model is about 87 MB, downloads in the background on first start, and hybrid search turns on at the next restart. If the download fails, search keeps working on full text.

  • Local, in-process (the default)
  • OpenAI
  • Voyage
  • Google Gemini
  • GitHub Copilot
  • Ollama, LM Studio, vLLM

Opt out with embedding_provider = "none". A provider you configured yourself is never overridden.

LLM: off by default

Add one when you want LLM-written consolidation, ai-memory bootstrap, richer linting, auto-improve or reranking. A small, fast model is the right size for this work, where the provider lets you pick one. With a Codex or ChatGPT login, keep the provider’s default model, because that backend rejects other model names. Subscription logins work, so a paid platform API key is not required.

  • Anthropic
  • Anthropic OAuth (Claude subscription tokens currently refused for third-party apps)
  • OpenAI
  • OpenAI OAuth / Codex CLI login
  • GitHub Copilot
  • Google Gemini
  • OpenCode Go and Zen
  • Ollama
  • LM Studio
  • vLLM
  • OpenRouter and any OpenAI-compatible endpoint

Check a provider with ai-memory llm-test. Fallback chains are supported.

Models, defaults and costs are in the LLM providers guide and the local embeddings guide.

Your agent is not listed?

Anything that speaks MCP can use the memory tools today.

Register a remote HTTP MCP server at this address in your agent’s settings. Search, page reads and writes, and handoffs all work. Automatic capture needs lifecycle hooks, so without them the agent only remembers what it writes.

Want full support for it? The managed-harness contribution guide describes what an adapter has to prove, and the contribute page says where issues and pull requests go.

MCP endpoint
# the default local server
http://127.0.0.1:49374/mcp

Questions and answers

Which coding agents does ai-memory work with?

27 agents and editors are listed, including Claude Code, Codex, Cursor, Gemini CLI, OpenCode. 20 of them have both automatic capture through lifecycle hooks and the memory tools over MCP.

What is the difference between hooks and MCP?

Hooks capture sessions automatically: prompts, tool calls and session boundaries are recorded without the agent doing anything. MCP gives the agent tools it can call to search and write memory. A fully supported agent has both.

My agent is not listed. Can I still use it?

Yes, if it speaks MCP. Point it at http://127.0.0.1:49374/mcp and it can search and write memory. Without lifecycle hooks, nothing is captured automatically.

Do I need an LLM or an embedding provider?

No. Both are optional. Since 2.0 the default embeddings run in-process with a model of about 87 MB that downloads on first start, and no API key is involved.

Does ai-memory run on Windows?

Yes. WSL2 is supported through the Linux path. Native Windows is experimental and ships as a release zip with ai-memory.exe.

Keep reading

Give your agents a memory today.

Free and open source under the MIT license. You need no account and no API key.

Install nowStar it on GitHub