ai-memory

Long-term memory for AI coding agents. Quit Claude Code mid-task, start OpenAI Codex in the same directory, continue without re-explaining the architecture, the failed approaches, or the open questions.

Why ai-memory

Your coding agent probably has a memory feature already. Claude Code takes its own notes, Cursor remembers some things, and every platform is adding more. They all have the same limits: the notes live on one machine, belong to one agent, and drop out of view when you switch tools or hand the work to a teammate.

ai-memory removes those limits.

• It follows you across agents. More than twenty harnesses (Claude Code, Codex, Cursor, Gemini CLI, OpenCode, Grok, Devin, Kimi, Kiro, and others) feed one shared memory. Quit Claude Code mid-task, open Codex in the same directory, and the next agent picks up a handoff that says where you left off, what failed, and what is still open. Handoffs are a typed protocol: each one has an owner and is claimed exactly once.

• It follows you across projects. A small profile of how you usually work (your package manager, test layout, architecture habits) reaches every new repository at session start, so you stop re-explaining yourself. It sits below each repository's rules file as a default. It is on by default for a single user, and on a shared server it is opt-in and private per person. See docs/cross-project-profile.md.

• It follows you across machines. Memory lives in a server you run, on the same laptop, a homelab box, or anywhere else, so the project you left on the desktop is the project you resume on the laptop, with the same knowledge and the same open questions.

• It works for a team. Point everyone at one server, and what one person's sessions learn, everyone's agents can retrieve. Knowledge is shared per project; personal handoffs stay personal. Multi-user auth, per-person attribution, and an audit log are built in and are not a paid tier.

• Your memory is plain markdown. The source of truth is a git-backed wiki of ordinary .md files: grep it, open it in Obsidian, edit it by hand, rsync it. The database is a derived index that can always be rebuilt from the files. There is no vector store to maintain and no content locked in a binary blob.

• It captures the work itself, silently. Lifecycle hooks record what happened (prompts, tool calls, session boundaries), sanitize it at a typed privacy boundary before anything is stored, and the server consolidates it into readable pages. You do not have to ask it to remember anything. The default path uses zero LLM calls: capture, search, and handoffs all work with no API key.

• It ages memory without an LLM. Memory decays on a schedule you can tune per tier, and memory you use decays slower: opening a page, searching it, or reaching it through a link counts as use. When episodic notes go cold they can be compacted down to their durable facts (file paths, error codes, decisions) instead of dropped. Near-duplicates collapse into one, and likely contradictions get flagged, all with zero API calls. Nothing is hard-deleted: the original stays in git and the version chain (restore-page brings it back). Access-weighted retention is always on because it can only keep memory longer. The parts that rewrite or drop content (compaction, dedup, per-tier curves) stay off by default until you turn them on.

• It can also "dream", if you let it. With an LLM configured, an opt-in background pass rewrites whole clusters of cold notes into single coherent pages while you are idle, and cancels as soon as you come back to work. It never deletes a source (the pre-merge versions stay reachable), it is off by default, and it has to pass a recall eval before it could become default behavior. Unless you ask for more, the zero-LLM path above is what runs.

• It is honest about itself. It ships as one self-contained binary. Purge commands say exactly what "deleted" means. The write ceiling (~700/s) is measured, not guessed. An audit log records every mutation.

How it works

capture ──▶ consolidate ──▶ recall ──▶ handoff
 hooks        session-end      search     next agent,
 observe      summaries as     + brief    any harness
 silently     wiki pages       injection

Agents emit sanitized observations through lifecycle hooks as you work. At session end, observations become coherent markdown pages in the project's wiki (optionally LLM-written; useful even without). The next session, in any agent on any machine, gets a bounded brief and can search everything: full-text, entities, links, and (optionally) vectors, fused into one ranking. Cross-agent handoffs pass the baton explicitly.

The full design, including the invariants that keep multi-user and multi-session use safe, is in docs/ARCHITECTURE.md.

Support matrix

Every row below is a first-party integration (MCP registration, lifecycle hooks, or both) checked by CI. The full matrix with per-agent notes and caveats is in docs/support-matrix.md.

Coming from another tool?

Most agent-memory tools optimize for one thing: extracting atomic facts per turn, a temporal knowledge graph, an agent-editable memory OS, or a hosted context API. ai-memory is built around a git-backed markdown wiki as the source of truth, with a derived index for retrieval. It captures automatically from lifecycle hooks, is shared across agents, machines, and people, and works with zero LLM calls by default. This table shows what carries over from each tool and what you gain:

Across all of these, ai-memory gives you files you own (git-backed markdown), a zero-LLM default, one self-contained binary, sharing across agents, machines, and teams, automatic lifecycle capture, and typed, claim-once handoffs. Opt-in features (LLM consolidation, vector search, the "dream" consolidation pass, belief-strength in ranking) stay opt-in, and the zero-LLM aging path (per-tier decay, extractive compaction, dedup, contradiction flagging, access-weighted retention) works with no API key at all.

Built on the shoulders of: the Karpathy LLM Wiki (compile-not-retrieve), agentmemory (this project is its Rust successor), basic-memory (markdown-on-disk truth), cognee (pipeline composition and triplet embeddings), Hermes Agent (the self-improvement loop), and A-MEM (Zettelkasten-style atomic notes).

The full comparison, covering where each approach wins, where ai-memory differs, and the published benchmark, is in How ai-memory compares.

Quick start

Arch Linux (AUR)

For native Arch installs, use the AUR packages. They install /usr/bin/ai-memory, packaged hook sources, and both system-level and user-level systemd units.

yay -S ai-memory-bin    # prebuilt Linux x86_64/aarch64 binary
yay -S ai-memory        # builds from source

Fedora (RPM)

Download the x86_64 or aarch64 RPM from the latest release, then install it:

sudo dnf install ./ai-memory-*.rpm

Then follow the native Linux service instructions in docs/install.md.

Single-user workstation:

mkdir -p ~/.config/ai-memory ~/.local/share/ai-memory
ai-memory --data-dir ~/.local/share/ai-memory \
  --config ~/.config/ai-memory/config.toml init
systemctl --user enable --now ai-memory.service
ai-memory install-mcp --client claude-code --apply
ai-memory install-hooks --agent claude-code --apply

System service installs use /var/lib/ai-memory and /etc/ai-memory/ via the packaged unit. Full user-service, system-service, auth, and provider setup is in docs/install.md#arch-linux-native-packages-aur.

macOS (menu bar app)

A self-contained .app that bundles the native ai-memory binary and hooks/ tree, starts the existing LaunchAgent, and opens /web, ai-memory status, and config.toml from the menu bar. Wiki, SQLite, config, and models stay in ~/Library/Application Support/ai-memory, so replacing the app is an update and does not rewrite that tree.

Needs a Rust toolchain and Xcode / Swift 6 (same as a source build):

git clone https://github.com/akitaonrails/ai-memory
cd ai-memory
./companions/ai-memory-macos/build.sh
open "companions/ai-memory-macos/dist/AI Memory.app"

Drag AI Memory.app to /Applications, then Install & Start Server from the menu extra (no Dock icon). When the status item is green, wire an agent with the bundled binary:

BIN="/Applications/AI Memory.app/Contents/Resources/runtime/ai-memory"
"$BIN" install-mcp --client claude-code --apply
"$BIN" install-hooks --agent claude-code --apply

Prebuilt tarball and launchd-without-the-app paths: docs/macos.md. Companion details: companions/ai-memory-macos.

Docker

You need: Docker or Podman + an agent CLI from the Support Matrix, or anything else that speaks MCP.

The published Docker image includes linux/amd64 and linux/arm64 variants, so Apple Silicon Macs and ARM64 Linux hosts can pull akitaonrails/ai-memory without --platform linux/amd64 emulation.

The default quick-start has no authentication. The server binds to loopback only, so on a single-user laptop nothing else can reach it. Adding a bearer token is a one-line change once you're ready to expose the server on the LAN; see Security below.

# 1. Install the ai-memory CLI wrapper (a small shell script that
#    runs the binary inside a container with your $HOME mounted). This is
#    the only thing that needs to live on the host filesystem.
mkdir -p ~/.local/bin
wrapper_tmp="$(mktemp -d)"
trap 'rm -rf "$wrapper_tmp"' EXIT
wrapper_base=https://github.com/akitaonrails/ai-memory/releases/latest/download/ai-memory-wrapper
curl -fsSL "$wrapper_base" -o "$wrapper_tmp/ai-memory-wrapper"
curl -fsSL "$wrapper_base.sha256" -o "$wrapper_tmp/ai-memory-wrapper.sha256"
expected="$(awk 'NR == 1 { print $1 }' "$wrapper_tmp/ai-memory-wrapper.sha256")"
if command -v sha256sum >/dev/null 2>&1; then
    actual="$(sha256sum "$wrapper_tmp/ai-memory-wrapper" | awk '{ print $1 }')"
else
    actual="$(shasum -a 256 "$wrapper_tmp/ai-memory-wrapper" | awk '{ print $1 }')"
fi
[ -n "$expected" ] && [ "$actual" = "$expected" ] || { echo "wrapper checksum mismatch" >&2; exit 1; }
install -m 0755 "$wrapper_tmp/ai-memory-wrapper" ~/.local/bin/ai-memory
rm -rf "$wrapper_tmp"
trap - EXIT
# Most distros put ~/.local/bin on PATH automatically. If `which
# ai-memory` comes up empty, add this to ~/.bashrc / ~/.zshrc:
#     export PATH="$HOME/.local/bin:$PATH"

# 2. Start the server. `--restart unless-stopped` makes it come back
#    on docker daemon restart and on machine boot (provided your
#    docker service is enabled at boot — `sudo systemctl enable
#    docker` on most distros). Loopback-only bind (`127.0.0.1:49374`)
#    so nothing outside this machine can reach it. Omit the LLM /
#    EMBEDDING lines for zero-LLM mode — FTS5 search still works
#    without any keys.
docker run -d --name ai-memory \
    --restart unless-stopped \
    -p 127.0.0.1:49374:49374 \
    -v ai-memory-data:/data \
    -e AI_MEMORY_LLM_PROVIDER=anthropic \
    -e ANTHROPIC_API_KEY=sk-ant-... \
    -e AI_MEMORY_EMBEDDING_PROVIDER=openai \
    -e OPENAI_API_KEY=sk-... \
    docker.io/akitaonrails/ai-memory:latest

# 3. Wire your agent CLI in two commands. The wrapper takes care of
#    mounts and each client's config-path detection. Re-run with
#    `--agent codex`, `--agent command-code`, `--agent devin`, `--agent opencode`, `--agent opencode2`, `--agent gemini-cli`,
#    `--agent grok`, `--agent kimi-code`, `--agent kiro-cli`, `--agent copilot-cli`, `--agent omp`,
#    `--agent oh-my-pi`, `--client cursor`,
#    `--client gemini-cli`, `--client grok`, `--client kiro-cli`, etc.
#    for additional agents; full list in docs/install.md.
ai-memory install-mcp   --client claude-code --apply
ai-memory install-hooks --agent  claude-code --apply

On Linux and macOS, the Docker wrapper runs install-hooks through its checksum-verified native host client. The installed hooks therefore enforce client-side capture controls such as [capture] ignore_paths and allowlist mode before an event reaches the spool or network. Set AI_MEMORY_HOOK_PLATFORM=posix explicitly only when you need the legacy shell compatibility path; the installer warns that path cannot enforce capture policy v1.

The examples use docker; replace it with podman on a Podman host. The wrapper automatically uses Podman when Docker is not installed. Set AI_MEMORY_DOCKER=podman to force Podman when both engines are available.

That completes the setup on Linux/macOS. Start a Claude Code session as usual: every prompt and tool call now lands in ai-memory, and the next session you open in this project will see a handoff with where you left off. On macOS the native binary is the recommended path when you do not need Docker, either through the menu bar app above or a release tarball / launchd agent. Later updates for that path use ai-memory upgrade (checksum-verified GitHub release replace + hook refresh); see docs/install.md#keeping-ai-memory-up-to-date. The same native upgrade path covers Windows x86_64 zip installs under a writable prefix (see docs/windows.md Scenario C).

Wiring another agent is the same two commands with a different name: --client codex, --agent codex, and so on for every row of the support matrix. OpenCode is version-detected by host-side commands; when generating its artifacts inside a container, use setup-agent --agent opencode --opencode-dialect v1|v2 --to /tmp/unused because the container cannot inspect the host executable. The opencode2 aliases remain force-V2 compatibility spellings. The full per-agent guide, including Windows and remote servers, is docs/install.md.

Two agents in the same project at once, or teammates on one server, work out of the box: the "current project" pointer is isolated per caller by default (v1.39+). See docs/auto-scope.md for the optional session-aware Claude Code bridge and the details.

If in doubt, start your harness with ai-memory run. It is the preferred way to launch: the first time it runs a harness it auto-installs that harness's ai-memory hooks + MCP if they are missing (so capture and recall work without a separate install-hooks/install-mcp step to forget), it wires the right project scope by construction, and it adds cross-harness session continuity on top of shared memory. Everything is idempotent and one-time per harness and config home. If the server is unreachable, run warns and launches anyway with local capture spooling (see Degraded offline launches); pass --require-server to fail closed instead.

ai-memory run claude
ai-memory run codex --yolo   # later: same workstream, different harness
ai-memory run --profile work claude  # reusable config.toml env/account preset
ai-memory continue           # resume the newest managed checkout
# after a dead launcher left its lease behind (same operator only)
ai-memory run --force-unlock codex
ai-memory resume             # pick from ALL workstreams in this checkout only
ai-memory resume --search auth # find a workstream by name (case-insensitive)
ai-memory resume --all       # pick across every linked checkout

--force-unlock immediately expires the selected workstream's active lease; use it only when you know the previous launcher is gone. It does not kill a native process, and it cannot evict another authenticated operator's run. See the managed-workstream recovery notes for the full safety contract.

In resume, type to search, use Up/Down to select a workstream, and Left/Right to choose its harness. Enter launches the selection; Escape clears a search, then cancels when the search is empty (Ctrl-C always cancels). The list scrolls and loads every checkout-local page; there is no default workstream cutoff. Use --limit N only when you want to cap the matching results. Workstreams from other repositories or worktrees are left out unless you pass --all, which lists every linked checkout (the current one first); continue still resumes the newest linked checkout from anywhere. Listing reads Git identity only; it does not scan the working tree with git status before showing the picker. Full checkpoints are still captured when launching the selected workstream.

Auto-wiring is on by default; opt out with ai-memory run --no-autowire or AI_MEMORY_RUN_AUTOWIRE=false. You can still wire agents by hand with install-hooks / install-mcp (e.g. for a harness you never launch through ai-memory run).

ai-memory uninstall --apply removes everything ai-memory installed, and only what it installed. It also clears ai-memory run's auto-wire record, so the next managed launch wires that harness again; to keep it unwired, launch with --no-autowire or set AI_MEMORY_RUN_AUTOWIRE=false. Install commands are idempotent and write timestamped backups next to any file they touch.

NixOS

This flake ships a NixOS module (nixosModules.default) with a systemd.services.ai-memory unit: a dedicated ai-memory system user (nologin, no linger) plus a hardened systemd sandbox (ProtectSystem = "strict", empty capability sets, MemoryDenyWriteExecute, RestrictAddressFamilies, and the rest; see nix/systemd-sandbox.nix). Packaged FHS units under packaging/systemd/ keep their existing lighter hardening. The flake exports packages for x86_64-linux, aarch64-linux, and aarch64-darwin. Intel macOS remains supported by the release tarball and Homebrew, but not by the pinned Nixpkgs revision.

{
  inputs.ai-memory.url = "github:akitaonrails/ai-memory";

  outputs = { nixpkgs, ai-memory, ... }: {
    nixosConfigurations.myhost = nixpkgs.lib.nixosSystem {
      system = "x86_64-linux";
      modules = [
        ai-memory.nixosModules.default
        {
          services.ai-memory = {
            enable = true;
            # enableWeb = true;  # off by default — the web UI is opt-in
            # enableApi = true;  # API-only companions; no browser UI
            settings = {
              allowed_hosts = [ "localhost" "127.0.0.1" "::1" "homelab.example" ];
              log_level = "info";
            };
            # Loopback (default): secrets optional; missing env file is tolerated.
            # Non-loopback: set one of ageSecret, sopsSecret, or environmentFile.
            # ageSecret = "ai-memory-env";  # config.age.secrets. (agenix)
            # sopsSecret = "ai-memory/env"; # config.sops.secrets. (sops-nix)
          };
        }
      ];
    };
  };
}

Declarative non-secret config lives in services.ai-memory.settings (a small typed set for common keys, plus freeformType for the rest of config.toml). The module renders a generated TOML file and passes --config. Top-level bind, port, and enableWeb win over duplicate settings keys. Anything in settings (including llm_headers) lands in a world-readable Nix store path, so do not put API keys there.

Secrets such as AI_MEMORY_AUTH_TOKEN never go in settings or the world-readable Nix store. This includes llm_headers, which can carry API credentials; set AI_MEMORY_LLM_HEADERS in ageSecret, sopsSecret, or environmentFile instead. These three secret sources are mutually exclusive. Non-loopback binds require one; loopback may omit them and tolerates a missing environment file (systemd EnvironmentFile=-…). Put TLS in front of a LAN/WAN bind; see docs/https-via-proxy.md.

All options and defaults are in nix/nixos-module.nix. The module is entirely opt-in: nix build, nix run, nix develop, and the CLI are unchanged if you don't import it.

Everyday use

Day to day, you mostly do not think about ai-memory. Hooks capture prompts, tool calls, and session boundaries; session end turns them into readable wiki pages; the next session starts with a handoff.

• Ask "where did we leave off?" to continue from the pending handoff.
• Ask "have we discussed X?" or "search memory for Y" to query the wiki.
• Ask "catch me up" for a prose digest of recent project activity.
• Run ai-memory bootstrap once when adopting an existing project with months of history.
• Start the server with --enable-web for a read-only browser view of the wiki and a JSON API under /api/v1.
• Back up host-side harness configuration with ai-memory backup-agents -o agent-assets.tar.gz; inspect a restore with ai-memory restore-agents -i agent-assets.tar.gz, then add --apply only after reviewing the active skills, plugins, instructions, and destinations.

Search modes, entities, feedback, briefings, and the web API are covered in docs/usage.md and docs/use-cases.md.

Teams and multiple machines

Run the server somewhere reachable, such as a homelab box or a LAN host, and point every machine and every teammate at it. Knowledge is shared per project; personal handoffs stay personal; every write is attributed and audited. Multi-user auth (passwords, API credentials) is built in.

Start with docs/users.md for accounts and ownership, and docs/deploy.md for the server itself. It includes measured capacity numbers and the one rule you must follow: one server per data directory, never two.

Security

The quick-start default is loopback-only with no auth, so nothing outside your machine can reach it. From there, hardening is incremental: a bearer token for the LAN, per-user accounts, OIDC device auth for hooks, TLS via a reverse proxy. Capture is sanitized at a typed privacy boundary before anything is stored, and per-repository [capture] rules can exclude paths or invert to allowlist mode. A repository can also route its capture to a different server than the one the hooks were installed against.

The full model is in docs/security.md, docs/users.md, and docs/https-via-proxy.md. For data-flow, identity/SSO, and offline-install questions specifically, see DATA_HANDLING.md, docs/sso.md, and docs/airgapped-install.md.

LLM providers

Optional. Everything works with zero LLM calls; adding a provider upgrades session summaries and enables semantic search. Anthropic, OpenAI (including OAuth), Codex CLI credential reuse, GitHub Copilot, Gemini, OpenCode (Go and Zen), and any OpenAI-compatible endpoint (Ollama, LM Studio, vLLM) are supported for consolidation; OpenAI, Voyage, Gemini, and keyless OpenAI-compatible endpoints for embeddings. Configuration lives in docs/llm-providers.md.

Architecture

One Rust binary runs an MCP/HTTP server and owns one data directory:

/
├── wiki/    # markdown source of truth, git-versioned
├── raw/     # immutable sanitized managed-workstream transcript segments
├── db/      # SQLite indexes, including FTS5, entities, and embeddings
├── models/  # reserved for local embedding models
└── logs/    # rolling tracing output

Hooks POST observations to the server. The server serializes writes through one SQLite writer, compiles session observations into markdown pages, and serves retrieval through FTS5, entity-match and graph-neighbor RRF, optional vector RRF, bounded source-authority adjustment, and bounded raw-observation fallback for non-global searches.

See docs/ARCHITECTURE.md for the data-flow diagram, crate breakdown, schema notes, and invariants.

Docs

For users

For contributors

Community articles

Written by users; not maintained here, so check them against the version you run.

• Installing and configuring ai-memory on Linux (pt-BR, LinuxPro, covers v2.6.3)
• ai-memory from v2.4 to v2.6.3 (pt-BR, LinuxPro)

License

MIT. See LICENSE.

Acknowledgements

This codebase is being built collaboratively with Claude Code (Anthropic Claude Opus 4.7) following the plan documented in docs/design-decisions.md.