Looks like your system
Linux
- x86_64ai-memory-linux-x86_64.tar.gz22.8 MBSHA-256 checksum
- ARM64ai-memory-linux-aarch64.tar.gz21.3 MBSHA-256 checksum
Unpack, run ./ai-memory init, then serve.
Developers
Download ai-memory v3.4.0
Released November 4, 2026. One self-contained binary per platform, each with a SHA-256 checksum next to it. MIT licensed.
Binaries
Looks like your system
macOS
- Apple Siliconai-memory-macos-aarch64.tar.gz20.6 MBSHA-256 checksum
- Intelai-memory-macos-x86_64.tar.gz22.2 MBSHA-256 checksum
Includes a launchd agent. If you download with a browser, clear the quarantine flag first.
Looks like your system
Windows
Experimental. WSL2 with the Linux build is the supported path. Scoop installs this same zip.
sha256sum -c ai-memory-linux-x86_64.tar.gz.sha256 # macOS: shasum -a 256 -c ...
After unpacking, follow the setup guide: install the binary, configure the server, then launch your agent.
Later, a binary under a writable prefix (Linux, macOS, the Windows zip) updates itself with ai-memory upgrade: it downloads the matching release, verifies the checksum and refreshes the hooks. Packages update through their package manager.
Coming from 2.x? Back up the wiki and database before upgrading the server, then update clients. Migrations are forward-only; rollback needs the old binary and backup. Read the 3.0 upgrade steps.
Package managers and containers
mise
mise use -g github:akitaonrails/ai-memory
Linux and macOS. Verifies the checksum for you. New releases can take about a day to appear.
Arch Linux (AUR)
yay -S ai-memory-bin # prebuilt
yay -S ai-memory # from source
Installs systemd units for a user or a system service.
Fedora (RPM)
sudo dnf install ./ai-memory-3.4.0-1.x86_64.rpm # or .aarch64.rpm
x86_64 and aarch64 packages with the binary, the hook sources and the systemd units. Then follow the user or system service steps.
Windows (Scoop)
scoop bucket add akitaonrails https://github.com/akitaonrails/scoop-bucket
scoop install ai-memory
Adds the maintainer's bucket, checks the zip against the checksum published with the release and puts ai-memory on PATH. Update with scoop update ai-memory; the built-in self-upgrade refuses a Scoop install.
Docker or Podman
docker pull akitaonrails/ai-memory:latest
Multi-arch: linux/amd64 and linux/arm64, so Apple Silicon needs no emulation.
From source
git clone https://github.com/akitaonrails/ai-memory
cd ai-memory && cargo build --release --workspace
Needs the Rust toolchain pinned in rust-toolchain.toml.
Client machines that only send hooks to a remote server do not need the full binary. The release also ships ai-memory-install-hooks and wrapper scripts for Docker on Linux, macOS and Windows. They are listed on the release page.
What changed
A short summary of the newest release, then the counts for the ones before it. The full changelog on GitHub has every entry.
New in v3.4.0
Your agent re-grounds after a compaction
When Claude Code, Codex or Factory Droid compacts a conversation, the hook injects a short anchor with the session's first and latest prompts instead of the usual startup packet. Every agent also gets a rule to re-read its own prompts after a compaction and trust them over the summary.
The interface speaks Portuguese
The browser and the human CLI output now ship English and Brazilian Portuguese. The browser follows Accept-Language and has its own selector; the CLI follows
--language,AI_MEMORY_LANGUAGEorlanguagein config.toml. Stored pages, JSON and logs are untouched.Install on Windows with Scoop
Add the maintainer's bucket and run
scoop install ai-memory. Each release pins the manifest to the Windows zip and the checksum published with it, and setup records the stablecurrentjunction instead of the version folder that cleanup deletes.A different model for each job
Consolidation, bootstrap, auto-improve, dream, lint, profile, reranking and query answers can each name their own provider and fallback chain. A task without one uses the default, and a task with one runs even when there is no default provider.
Bring a claude-mem database in
The optional importer reads supported claude-mem SQLite schemas from a read-only snapshot, with explicit project mapping, client-side scrubbing and a dry run by default. Behind it, a guarded batch write lands 1 to 100 pages in one project atomically.
Two retrieval signals you can turn on
One lowers the authority of a page carrying open stale or wrong feedback until a rewrite supersedes it. The other boosts a page a session read or marked helpful for a similar query. Both weigh 0.0 by default, and both show up under
explain=true.Smaller switches
Index the path of the file a tool touched, leave no handoff for one run with
AI_MEMORY_AUTO_HANDOFF=off, turn the recovery snapshot off per project, see pending mail on a prompt, and read up to 50 recent sessions in a briefing. All of them are off by default.
Also in this release:25 added2 changed4 fixedChangelog entry
Earlier releases
Give your agents a memory today.
Free and open source under the MIT license. The default memory features need no model API key; your coding agent may require its own account.