Semantic Memory — AI Glossary

Reference details

Related terms

• Agentic Memory
• Knowledge Graph
• Memory (Agent)

How it works

1. Candidate facts are extracted from conversation or documents and written as standalone statements, since a fact that only makes sense in its original context cannot be retrieved usefully.
2. Each is stored with provenance — when it was learned and from what — because the entry is a model's claim rather than a verified record.
3. Retrieval is by relevance to the current turn, so only a few facts enter the prompt instead of the whole store.
4. Contradictions are reconciled on write: a new fact that conflicts with an existing one must supersede it, or the store will return both and the model will pick arbitrarily.
5. Entries are re-validated or expired, because a fact extracted once and never checked becomes stale silently and is retrieved with the same confidence.

Example: Why a contradiction is worse than a gap

An agent learns in March that a user's preferred region is EU, and in September that it is US. Stored without reconciliation, both entries match a region query and the model selects one — in practice often the better-embedded phrasing rather than the more recent fact. The user then sees a confident answer based on a six-month-old preference. An empty store would have prompted a question; the unreconciled one produced a wrong answer with no signal that anything was uncertain.

Commonly confused with

The counterpart to episodic memory: generalised facts rather than a record of events. It differs from a knowledge base in provenance — a knowledge base is authored and reviewed, whereas these facts were extracted by a model from conversation and inherit that model's errors. Treat entries as claims with a source, not as ground truth.

When to use it

Reach for it when:

• Durable facts about a user or domain should persist across sessions without being re-stated.
• Re-stuffing the same background into every prompt is wasting context.
• Facts are reusable across tasks rather than tied to one episode.

Reach for something else when:

• The information must be authoritative, where an authored and reviewed knowledge base belongs instead.
• The facts are episodic — what happened, and when — which is a different store.
• Nothing reconciles contradictions, in which case the store will eventually answer confidently and wrongly.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Semantic Memory — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path