Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

ai-memory

Fabio Akita's self-hosted memory server: hooks record agent sessions into a git-backed Markdown wiki shared by 20+ coding agents.

Reviewed by ReviewRamble editors · Updated

View on GitHub Open source

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

4.2/5

ai-memory is the agent memory to run if you switch between Claude Code and Codex, because the handoff, not the recall, is what it does better than anyone else.

What is ai-memory?

ai-memory gives coding agents one memory that outlives the session, the agent and the machine. Lifecycle hooks record prompts, tool calls and session boundaries, strip secrets at a privacy layer, and post them to a local server; at session end the server turns them into Markdown pages in a git-versioned wiki. The next session, in any supported agent, starts with a short brief and a handoff saying where work stopped, and can search the wiki over MCP.

The server is one Rust binary with SQLite full-text search, entity and link ranking and optional vectors. The default path makes no LLM calls; an Anthropic, OpenAI, Gemini or local model can be added to write better summaries, and an opt-in background pass merges cold notes. It supports more than 20 agents, including Claude Code, Codex, Cursor, Gemini CLI, OpenCode, Kiro and GitHub Copilot CLI, and a team can share one server with per-user accounts and an audit log.

It is written by Fabio Akita under the MIT licence. The repository was created in May 2026; v3.5.1 shipped on 19 November 2026.

Is ai-memory worth it?

Most memory add-ons help one agent remember; ai-memory is built for the moment you quit one agent and open another. A handoff is a typed record with an owner that the next session claims once, and the ai-memory run launcher goes further: start Codex after Claude Code hits its limit and Codex gets the shared workstream plus a temporary local copy of Claude's visible history, which is deleted when it exits. The same server can serve several machines and a whole team, with knowledge shared per project and handoffs kept per person.

The design choices are conservative in a useful way. The source of truth is plain Markdown in git that you can grep, edit or open in Obsidian; the SQLite index can be rebuilt from it. Capture, search and handoffs need no API key. On the project's own LongMemEval-S harness, the zero-LLM search scored a hit@5 of 0.953 for 3.5.0 and local embeddings 0.940, with every run's commit and hardware published, though no one outside the project has reproduced them that we found. Aging, compaction and dedup run without a model and never hard-delete; the parts that rewrite content are off by default.

The security notes are unusually complete. A native install binds to loopback and mints an owner-only bearer token; the Docker quick-start does not, and runs without auth until you enable it. There is no telemetry and no update check, but the first start downloads about 87 MB of embedding model from Hugging Face unless you turn local embeddings off, and nothing is encrypted at rest. Remember what the hooks do: every prompt and tool result lands in the store, after sanitising, so a shared server holds a lot of your work.

The weight is the cost. The README alone runs to 47 KB, the changelog to 400 KB, and releases land every few days; the author credits Claude Code with writing much of it. Native Windows is experimental, writes top out near 700 a second, and only one server may own a data directory. If you use a single agent on one laptop, Claude Code's own memory is less to run.

Where it earns its place

  • A freelance developer who runs out of Claude Code quota mid-afternoon

    Carry on in Codex without re-explaining what was tried and what failed

    Launch both through ai-memory run; when Claude stops, ai-memory run codex picks up the handoff and a temporary copy of Claude's visible history. Free; no API key needed for the default path.

  • A four-person team sharing a homelab box

    Let each person's agent find decisions a teammate's sessions already made

    Run one server on the LAN behind a TLS reverse proxy, create user accounts, and point every machine's hooks at it. Knowledge is shared per project, handoffs stay personal, and every write is attributed in the audit log.

  • A developer at a firm that needs a data-protection review before installing tools

    Show exactly what leaves the laptop

    Hand over DATA_HANDLING.md, which lists the one default download and five opt-in paths that send data out, and set embedding_provider to none for no outbound calls at all.

Before you commit

Setup
curl the install script (it checks the SHA-256 before installing) and run ai-memory setup, which creates the data directory and token, starts a user service and wires MCP, hooks and skills for each agent it finds after asking. Homebrew, AUR, RPM, Scoop, Nix, Docker and a macOS menu bar app are also offered.
Integrations
MCP plus lifecycle hooks for Claude Code, Codex, Cursor, Gemini CLI, OpenCode, Kiro, GitHub Copilot CLI, Factory Droid and others; marketplace plugins for Claude Code and Codex; a JSON API and read-only web view; a TypeScript client and a document importer for PDF, DOCX and transcripts.
Data and privacy
Everything lives in one local data directory: a git-versioned wiki and SQLite indexes. No telemetry or update check. Local embeddings download about 87 MB from Hugging Face on first start. Optional cloud LLM or embedding providers receive page text. No encryption at rest.
Cost at scale
Free to run; the default needs no model. Turning on a cloud LLM for summaries, an embedding API or the background merge pass adds provider charges per session, which the project does not estimate.
Limits
About 700 writes a second; one server per data directory; native Windows experimental; human logins beyond loopback need HTTPS unless you explicitly allow plaintext.

What’s good

  • Claim-once handoffs and session carry-over across 20+ agents
  • Plain Markdown in git is the source of truth, not a database
  • Capture, search and aging work with zero LLM calls
  • No telemetry; loopback bind and bearer token by default

What’s not

  • Large, fast-moving codebase with releases every few days
  • Docker quick-start runs without authentication
  • Benchmark figures come from the project's own harness
  • No encryption at rest; native Windows is experimental

Best for

Developers and small teams who move work between Claude Code, Codex and other agent CLIs, or between machines, and want the shared memory in files they own.

ai-memory alternatives

ai-memory: common questions

Is ai-memory worth it?
ai-memory is the agent memory to run if you switch between Claude Code and Codex, because the handoff, not the recall, is what it does better than anyone else. We rate it 4.2 out of 5.
How much does ai-memory cost?
Free under the MIT licence; optional LLM and embedding providers are billed by those providers.
What are the downsides of ai-memory?
Large, fast-moving codebase with releases every few days; docker quick-start runs without authentication; benchmark figures come from the project's own harness; no encryption at rest; native Windows is experimental.
Who is ai-memory best for?
Developers and small teams who move work between Claude Code, Codex and other agent CLIs, or between machines, and want the shared memory in files they own.
How much does ai-memory cost for a team or in production?
Free to run; the default needs no model. Turning on a cloud LLM for summaries, an embedding API or the background merge pass adds provider charges per session, which the project does not estimate.
Is ai-memory safe for company data?
Everything lives in one local data directory: a git-versioned wiki and SQLite indexes. No telemetry or update check. Local embeddings download about 87 MB from Hugging Face on first start. Optional cloud LLM or embedding providers receive page text. No encryption at rest.
What are the best ai-memory alternatives?
The closest alternatives we have reviewed are claude-mem, Hindsight and Mem0.

Answers checked Nov 20, 2026.

Make ai-memory? Show this rating on your siteai-memory rated 4.2 out of 5 on ReviewRamble