One memory you own. Every AI you use.
Stop repeating project context, preferences, and decisions to every assistant. Memoars gives your AI tools the same approved memory while keeping you in control of what is learned, changed, and forgotten.
Follow a single memory, end to end.
From the moment your agent writes something to the moment another agent recalls it. Memory content leaves your machine as ciphertext; operational metadata needed for coordination remains visible to the service.
Your key never crosses into the shaded zone. Memory content is ciphertext, while organization, workspace, identity, grant, version, usage, transcript, and proposal metadata may be visible to the coordinator.
And then it helps you curate.
COMING SOONOn an opt-in run, the dreamer will review this lifecycle and propose what to keep, update or merge. You approve every change. Planned — not yet part of the product.
Local works alone. The coordinator is how a memory becomes shared.
Running memoars against your own bucket is free and complete: one agent, one machine, fully encrypted. Once a second agent, a teammate, or a second device needs the same memory, someone has to settle whose write wins, who may read it, and which version is current.
A shared bucket alone can't do that safely. Two agents racing to write will clobber each other, and access control becomes an honour system. The coordinator referees concurrent, multi-party memory while never being able to read any of it.
Concurrency that can't clobber
Compare-and-swap on version + hash. If two writers collide, the conflict reconciles and retries instead of one silently overwriting the other.
Permissions that are cryptographic
Orgs → workspaces → identities, with per-workspace read/write/propose/approve grants. Each workspace has its own passphrase, so workspace isolation is enforced by encryption as well as by the API.
Cross-device sync + attribution
Every device stays current, every change is attributed to its writer, and an out-of-band dreamer can propose updates, with a human approving before anything lands.
Review AI-proposed memory changes before they land.
Left alone, memory only grows by explicit writes and slowly rots as duplicates pile up and facts go stale. Dreaming proposes the cleanup — you approve it.
When you run it, a key-holding dreamer reviews a bounded set of your agents' transcripts against existing memory and works out what's worth keeping: new facts to add, entries to update, near-duplicates to merge. It never edits memory directly. It queues proposals for you to review and approve. Remote review is explicit opt-in, and this capability is not yet part of the product.
Stale entries get refreshed. Facts that changed are flagged and updated instead of quietly misleading your agents.
You review, not groom by hand. Instead of curating memory manually, you approve or dismiss a short list of suggested changes.
Guarded against memory poisoning. A human approves every proposal before it is applied. Dreamer deployments can access transcript and proposal content and must be trusted accordingly.
Bring your own storage.
Memory content is encrypted client-side before upload, so even a public-read bucket exposes only encrypted memory content.
Contact
Memoars is currently an invite-only beta. Request beta access here, ask a question, or tell us about an enterprise or self-hosted deployment — we read every message.