I struggled a bit with context sharing, knowledge sharing, memories sharing between AI agents (I use two or three on a daily basis). Each of them has its own memory, they dont share it or its a bit cumbersome to do memory curation and improve it (especially if there are some API AI calls that run occasionally from different models)Memoars is an attempt to solve it - one memory that belongs to you (no storage vendor lock) that any assistant can read and write through MCP.How it works:- Memory content is encrypted on your machine (XChaCha20-Poly1305, key derived with Argon2id) and written directly to storage you own - R2, S3, MinIO, Supabase etc)- A coordinator handles the metadata plane: sequence numbers, versions, grants, conflict resolution. It never receives the workspace content key, so it can t read memory content. It does see operational metadata - org, workspace, identity, version, usage- Every change lands in an append-only, hash-chained log with compare-and-swap on writes, so two clients can t silently clobber each other and you can see how a memory got to its current state.- Permissions are orgs → workspaces → identities, with per-workspace grants. Each workspace has its own passphrase, so isolation is enforced by encryption as well as by the API.Where it actually is:It s invite-only right now, and I want to be honest that this is a invite list rather than a product you can go install this afternoon (as I want to make sure it makes sense and that it solves a problem for you before its shipped).The client is being open-sourced and the hosted coordinator opens shortly after.Drop me a mail or ping me for more info - as I will reply to all inquiriesTnx for taking a look!