The deterministic storage backbone.
LatticeCore is the data layer that stores and retrieves organizational context through three memory tiers: P1 CanonStore for exact-match facts, P2 LatticeMem for session and role-aware memory, and P3 Drop Corpus for semantic fallback. Every retrieval carries an OTel-correlated trace ID.
A single vector store is not a knowledge system.
Most AI systems store context in a single embedding store and retrieve by semantic similarity. This works for narrative questions but fails for factual ones. "Who owns the billing service?" requires an exact-match lookup, not a similarity search. "What are the deployment constraints for production?" needs role-scoped access, not the nearest embedding. Semantic approximation produces confidently wrong answers for relationship and ownership queries.
LatticeCore solves this with tiered memory. P1 CanonStore handles deterministic fact retrieval through exact-match quads. P2 LatticeMem delivers session and role-aware aggregates with permission scoping. P3 Drop Corpus provides semantic fallback as a last resort. Tier precedence is enforced in code — deterministic answers always supersede semantic approximations.
Deterministic first. Semantic last.
Every context request routes through three tiers in strict precedence order. The right tier answers the right kind of question.
CanonStore
Exact-match quads for verified facts and relationships. SPARQL queries return deterministic ownership, dependencies, and source-of-truth precedence. The independent safety anchor for the entire platform.
LatticeMem
Session and role-aware memory with permission-scoped retrieval. Holds curated atoms, employee contributions, decisions, and workflow scenarios. Graph-RAG over the company context graph.
Drop Corpus
Vector search over raw source material. Only consulted when deterministic tiers produce no match. Returns narrative context, never used as the source for ownership, status, or precedence answers.
The storage backbone every other product depends on.
Semantic similarity is not factual accuracy. A single vector store queried by embedding will return semantically proximate but factually incorrect answers for relationship and ownership queries. LatticeCore enforces deterministic tier precedence so that P1 CanonStore answers always supersede P2 aggregates, which always supersede P3 semantic results.