What Mnemosyne is, and what actually runs today

A tamper-evident record of every ingest, query, retrieval and answer — change one sealed block and the whole chain says so.

The chain's root is published into Bitcoin, so a root already there cannot be un-published — even by someone who owns the box and its keys.

The block-sealing key can rotate without invalidating any history, and no key can be spliced in without the outgoing key's signed consent.

Browse the tamper-evident chain yourself — every block, every anchor, no key required, sealed content stays sealed.

Access is a lattice, not a flat list — a reader must out-rank the level, hold every compartment, and clear every caveat, and it is checked when the key is wrapped.

Who may install an access-policy change is decided where the chain is read, not only where it is written — so no back-door write can grant itself the encryption gate.

Every clearance grant, revocation, and key event is a signed entry on an anchored log — so who could read what, and since when, is answered by replaying the chain, not by trusting a config file.

The keys that unlock everything are never written to disk in the clear — a stolen disk yields wrapped keys with nothing to unwrap them.

Data at rest is encrypted under a key scoped to its full label-set — with two known gaps we name openly, both closed before any real customer data is ingested.

Identifiers that leak nothing — a stolen id is a hash, decodable only by a cleared operator holding the key.

Blocking not just direct disclosure but inference toward a restricted fact — a three-check frame gate plus a human-ratification ratchet.

Fool the AI, and your data is still safe — the model is untrusted by design, so a compromised model still cannot invent a fact or leak one across compartments.

Every claim in an answer must resolve to a source the requester is cleared to see — and that check runs outside the model, so fooling the model does not move it.

We curate your sources once into a verified, interlinked graph that compounds — unlike retrieval that re-derives a raw-chunk answer from scratch on every query.

Retrieval is ranked by a cognitive model of memory — spreading activation over the knowledge graph — not bare vector similarity.

The system gets measurably better with use — and learning can only tune ranking and action-selection, never the security gate.

Standard company files are parsed, mandatorily clearance-tagged, and sealed on the way in — labels are final before the first byte is anchored.

A company brain you run on your own infrastructure — no data leaves your building, and an air-gapped deployment has nowhere for a leak to go.

A private, single-tenant instance we run for you — one customer per box, never shared, and you can leave with your data any time.

We run your instance and still cannot read it — your keys unlock it only inside an attested enclave the host cannot see into.