Semantic Memory

What it is

The agent's store of standing claims about the entities it deals with — the person it is working for, their accounts, their stated preferences, the systems it operates against. Each item is asserted as currently true, carries no occasion, and is expected to be replaced when it stops being true: deploys go out on Thursdays, the finance contact is Priya, the preferred units are metric. These are beliefs the agent formed, which is what separates them from the documents an organisation curates and publishes.

Why it matters

This is the tier that makes an agent appear to know someone, and the tier that makes it confidently wrong about them. Every item in it is something the product will later assert without asking again, so the bar for writing one is a product and trust decision rather than a storage decision. A belief store without a revision path accumulates claims that were true once and are indistinguishable at retrieval time from claims that are true now, and the agent has no way to hedge because nothing told it which is which.

Key decisions

• Whether beliefs are extracted automatically from dialogue or written only on explicit confirmation. Extraction scales and it manufactures claims out of hypotheticals, jokes, quoted third parties, and things the user was arguing against.
• The unit: a typed subject-predicate-value slot, or a free sentence. Slots make replacement mechanical and reject what does not fit; sentences accept everything and turn replacement into a search problem.
• Whether an item carries a confidence and a learned-at date, and whether anything downstream actually consults them rather than storing them decoratively.
• Scope. The same sentence can be a reasonable belief about one person and a leak when held at account or organisation level, and the scope is fixed at write time.
• How the agent is permitted to use a belief in output: stated as fact, hedged, or attributed to when and how it was learned. This is the difference between a product that feels attentive and one that feels presumptuous.
• Whether the person the beliefs are about can read and edit them, which is the only correction mechanism that does not depend on the system noticing its own error.

Failure modes

• A user says "suppose I lived in Spain" and the extractor writes that they live in Spain.
• A preference stated once for one task — "keep it short" — is stored as a standing instruction, and long answers are never produced again even when explicitly requested.
• Items are written with no learned-at date, so nothing can decay, nothing can be ranked by freshness, and nothing can be audited to a time window.
• Writes overwrite in place, so a wrong extraction is unrecoverable and the user cannot be shown what changed about them or when.
• A belief inferred from one customer's conversation is keyed at organisation scope and surfaces in a different customer's session.
• The store holds two incompatible claims about the same subject with no supersession, and the answer depends on which one the index returns first.
• Beliefs are injected into prompts unlabelled, so the model treats a year-old guess with the same authority as something the user said thirty seconds ago.
• An extraction prompt is changed and the quality of newly written beliefs shifts, but nothing records the extractor version, so the affected cohort cannot be identified afterwards.

Related capabilities

• Episodic Memory — the occasions that are evidence for these claims
• Memory Conflict Resolution — what decides between incompatible claims about one subject
• Memory Provenance — whether a claim was stated, observed, or inferred
• Long-term Memory — the durable tier these are stored in
• Memory Retrieval — how a claim reaches the model, and with what label

Applies: Production readiness · Design questions