What is Semantic Memory?
Also called Fact Memory, Factual Memory.
Semantic memory is an AI system's store of general facts and concepts detached from the occasion on which they were learned, such as a customer's billing cycle or a company's approval threshold. It answers what is true rather than what happened, and it is typically held as structured statements, graph edges, or indexed documents.
The distinction from episodic memory is the presence or absence of a specific occasion. Semantic entries are timeless statements: this account is on an annual plan, this region requires two approvals. They can be derived from a single conversation or distilled from many episodes, and once stored they are recalled without reference to their origin event.
Because entries are decontextualized they are compact and reusable, which makes them the most efficient form of long term recall. That same property makes them dangerous when they go stale, since a statement stripped of its occasion carries no cue that it might have expired. Timestamps, sources, and confidence markers mitigate this risk.
Representation varies. Some systems store natural language statements and retrieve them by embedding similarity, others store subject relationship object triples in a graph, and many combine both so statements remain readable while relationships remain traversable. The graph form supports aggregation and multi hop queries that free text statements do not.
Maintenance is the recurring cost. New information can contradict an existing entry, so systems need consolidation routines that merge duplicates, supersede outdated statements, and resolve conflicts. Without them the store accumulates near duplicates that compete during retrieval and produce inconsistent answers depending on which copy surfaces first.
Key points
- General facts detached from when they were learned.
- Compact and reusable, which makes recall efficient.
- Stale entries carry no cue that they expired.
- Stored as statements, graph triples, or both.
- Requires consolidation to merge and supersede duplicates.
In practice
Across several conversations a user mentions working in the European Union, invoicing in euros, and closing books quarterly. Those three statements are extracted into semantic memory with sources and dates. Months later, when the user asks for an invoice template, the system applies the euro currency and the quarterly period without needing the original conversations.