Marciana is the memory and cognition product layer in the QueryGraph stack. It composes TypeSec, Grust, Sail, and LakeCat-governed catalog integration behind a native four-verb API:
rememberingests protected memory;recallretrieves through the TypeSec vault;improveruns durable cognition and applies inert proposals through a guarded commit; andforgetperforms scoped, audited removal or retraction.
The repository is an extracted, verified baseline. The querygraph-memory
transplant preserved behavior, durable identifiers, and tests, and the
workspace is now split into four product and adapter crates:
marciana-ledger, marciana-cognition, marciana-memory, and
marciana-catalog.
TypeSec remains the authority for capability checks, labels, retention,
quarantine, protected-content access, proposal validation, and TypeDID
verification. Grust owns generic graph and transaction mechanics. Sail owns
generic distributed execution. LakeCat owns governed-scan proofs. Marciana
owns memory-product orchestration and adapters, and QueryGraph consumes it.
LakeCat is required for catalog-backed governed improve; local memory use may
run the native lifecycle without a catalog adapter.
Only the capability-gated TypeSec vault may reveal or mutate protected memory. Cognition proposes, indexes rank, and stores persist; none of them is an authorization authority.
Cognee is inspiration only, not a compatibility target. Marciana's baseline is complete against its native four verbs and does not depend on Cognee's runtime, adapters, stores, API shape, or completeness model. A future edge adapter would require a separate decision and could only lower into the native contract.
- DESIGN.md is authoritative for project ownership, dependency direction, trust boundaries, and Sail integration policy.
- COMPATIBILITY.md records the cross-stack compatibility contract and points to machine-readable revision pins.
- CHANGELOG.md records every user-visible logical change.
- AGENTS.md defines contribution and delivery rules.
The workspace can be inspected with:
cargo metadata --no-deps --format-version 1CI requires formatting, lint, workspace tests, conformance tests, the dependency-free Python benchmark suite, and a live Sail gate built from the exact recorded upstream revision.