Canonical identity
Durable entities represent people, organizations, products, projects, services, and other promoted subjects. Alias and merge decisions avoid treating a shared name as sufficient identity proof.
Memory quality
Canonical identity, provenance, observation time, uncertainty and conflict keep durable context useful without presenting it as certainty.
Implemented model
The canonical ledger holds distinctions that a flat document or vector result cannot reliably express on its own.
Durable entities represent people, organizations, products, projects, services, and other promoted subjects. Alias and merge decisions avoid treating a shared name as sufficient identity proof.
Published memory retains bounded source references and lineage back to the contribution that supports it. Detailed private evidence remains outside routine retrieval.
Claims can carry when something was observed and whether it is current, historical, or superseded. Freshness is not inferred from retrieval rank alone.
Contradictory evidence can remain represented as conflict or uncertainty. The system is designed not to flatten disagreement into one confident profile.
Operators can correct identity and memory state, and successful source reprocessing atomically replaces the previous contribution from that source.
Memory cards are derived for a named platform, subject, agent, and scope. A card is a bounded projection, not a universal or psychologically complete profile.
Temporal and conflict states
A useful answer may need current context, but the memory system must retain enough state to explain what changed and where certainty stops.
The best supported context intended for present use, subject to its evidence and observation time.
Previously true or relevant context retained for chronology rather than presented as current.
A claim or source generation displaced by a newer active contribution without erasing its lineage.
A bounded interpretation derived from evidence, kept distinct from an explicit fact or direction.
Material sources disagree and the system preserves the disagreement instead of selecting certainty silently.
Evidence or identity resolution is not strong enough for an unqualified memory statement.
Provenance and publication
Private source contributions preserve source identity, generation, and derivation details. Published pages and scoped memory cards retain bounded citations and quality cues suitable for inspection without exposing the whole private ledger in routine retrieval.
Each processing run can be tracked as staging, active, superseded, or failed, so publication can point to the generation that contributed the memory.
Retrieval surfaces can expose bounded source references, observation time, and qualification without publishing raw private evidence by default.
Ambiguous identity, merge proposals, conflicts, and corrections can require an inspectable human decision rather than automatic resolution.
Current limitations
The product is designed to make error states visible and reversible. It does not eliminate hallucinations, infer hidden mental states, or create complete knowledge from sparse signals.
Names, email fragments, and contextual mentions may be insufficient to merge records safely. The correct outcome can be separate or unresolved identities.
Preserving disagreement is sometimes more accurate than selecting one source. A downstream agent must respect that qualification.
The ledger and operator controls support correction and provenance inspection; the complete self-service customer correction experience is not yet a general product claim.
Evaluate the memory
Strong beta partners review identity, sources, freshness, conflicts, and the retrieved context—not only whether the final answer sounds plausible.