Your agents will change. Their memory shouldn't have to.
OmniMem is open source, self-hosted memory for AI agents. Claude, Codex, Cursor, Copilot, Kiro, Lumo and anything else that speaks MCP share one store of what your teams have learnt, on hardware you own. Sponsoring it keeps that memory independent of whichever model vendor is ahead this quarter.
- MITlicensed, self-hosted
- 12agents with a setup guide
- −26%tokens per task, measured
- 18,000+downloads since March
No lock-in
Memory is the switching cost every vendor is building. Keep it in an open server and you can change agents without losing what they learnt.
Fewer tokens
An agent that already knows the project, and what failed last time, does less rediscovering. In our runs that was a quarter fewer tokens.
Sovereign by default
What an assistant remembers is as sensitive as the chats themselves. OmniMem keeps it on your machines, with embeddings run locally.
Built on MCP, so it's never tied to one vendor
The Model Context Protocol is an open standard, and every major agent now speaks it. OmniMem is an MCP server, which means the same memory serves Claude Code on one desk, Codex on the next and Copilot in the IDE down the corridor. Nobody has to agree on a single assistant for the knowledge to be shared.
That matters more every month. Vendors are adding memory of their own, and it's good, but it lives in their cloud and leaves when you do. Swap Claude for Codex, or Codex for whatever ships next spring, and OmniMem's memory comes along untouched. A full backup is a plain JSON file.
- Mixed teams share decisions, fixes and dead ends across agents
- Trial a new agent against the same memory and compare like with like
- OAuth 2.1 for browser clients like claude.ai, Le Chat and Lumo
- Hooks in Claude Code, Codex, OpenCode, Mistral Vibe and Kiro that brief every session
on your hardware
Fewer tokens, smaller bills, less waiting
We gave Claude Code the same task twenty times on a fresh copy of the code, ten runs without OmniMem and ten with it. The task's TODO recommends a crate that fails at runtime. With OmniMem the earlier failure is already in memory, so the agent skips the dead end and gets on with the job.
Tokens
−26%
833k → 616k median tokens per run
Cost
−26%
$2.16 → $1.51 median per run
Time
−35%
median wall time per run
Dead ends
0 of 10
runs repeated the known failure, against 8 of 10
Where the saving comes from
- A briefing, not a re-explanation
- Each session starts with what the project already knows, so the agent doesn't spend its first ten minutes reading the codebase to rediscover it.
- Dead ends answered back
- An approach that was tried and abandoned comes back as a warning, with the reason and what worked instead, before the agent burns a test run on it.
- Only what's relevant
- Recall drops anything below a relevance floor rather than padding the context, so an empty answer costs a few tokens, not a few thousand.
- Search that costs nothing
- Embeddings run on your own machine, so finding a memory never touches a paid API.
OmniMem 7.0.12, Claude Haiku 4.5, one scenario and ten pairs, so treat the token saving as promising rather than proven (a sign test gives p = 0.05). A dead end that takes longer to find would be worth more, and across a team the same lesson is saved once and used by everyone. Read the full report (PDF).
What sponsorship pays for
OmniMem is built and maintained by SquareCows Ltd, a small independent company. Sponsorship goes on the work that makes it safe for a company to adopt, and the things a small team can't pay for on its own.
A signed Windows installer
The MSI isn't signed yet, so SmartScreen warns on every install. A code-signing certificate fixes that for every user.
Independent benchmarks
More scenarios, more models and more agents than the ten pairs above, run so anyone can repeat them from the open test harness.
A security review
An outside audit of the OAuth server, the settings panel's sign-in and the audit log, published in full.
Every new agent, quickly
A setup guide, and hooks where the agent allows them, within weeks of a new client shipping MCP support.
Flathub and the app stores
Offline, reproducible builds for Flathub, and the packaging work to keep macOS, Windows and Linux releases in step.
Maintainer time
Fixing bugs, reviewing contributions and answering issues, so the project doesn't depend on someone's evenings.
Pick a level that suits you
Backer
$100 a month
For teams already using OmniMem who want it to stay healthy.
- A vote on which features come next
- Your logo on this page
- A thank-you in the release notes
- Sponsor label on your issues
Partner
$500 a month
For companies rolling it out across several teams or agents.
- Everything in Backer, logo on the homepage
- Help with your first deployment
- Your issues looked at first
Founding partner
$2,500 a month
For organisations that want an open memory layer to exist, long term.
- Everything in Partner
- A quarterly call on the public roadmap
- A named benchmark or guide for your stack
- A say in which agents and platforms come next
Sponsoring on your own? The Supporter tier is $5 a month, and any amount, monthly or one-time, is welcome on GitHub Sponsors. Every sponsor gets a vote on which features come next. Each vote opens in the sponsors' newsletter on GitHub, with the shortlist and how to vote.
What sponsorship doesn't buy
The self-hosted OmniMem stays MIT licensed, with no features held back for sponsors. The roadmap stays public and the code stays on an open forge. Sponsors get heard first, not obeyed.
Be among the first
If you want secure, sovereign AI, don't 'forget' about its memory.
Help keep the memory layer open, self-hosted and on speaking terms with every agent.