Semantic Memory

System retrieval published

Also known as: Factual Memory, Knowledge Memory

Definition

Memory of facts, concepts, and general knowledge independent of specific episodes or events. In AI agents, semantic memory stores learned facts, user preferences, and domain knowledge that can be retrieved and applied across contexts. Unlike episodic memory (what happened), semantic memory captures (what is true).

What this is NOT

  • Not episodic memory (episodic is events; semantic is facts)
  • Not the model's parametric knowledge (that's in weights; semantic memory is external)
  • Not a knowledge base for RAG (though it can use similar technology)

Alternative Interpretations

Different communities use this term differently:

llm-practitioners

A structured or semi-structured store of facts and knowledge that an agent can query. May be implemented as a knowledge graph, key-value store, or structured entries in a vector database with metadata.

Sources: MemGPT semantic memory design, Knowledge graph integration patterns, Entity memory in LangChain

cognitive-science

The memory system for general world knowledge—facts, concepts, and meanings—independent of the context in which they were learned. Distinguished from episodic memory (personal experiences).

Sources: Tulving: Semantic memory (1972), Cognitive psychology literature

Examples

  • Agent knows 'User's name is Alice' without recalling when this was learned
  • Storing entity information: 'Project X uses React, started in 2023'
  • Accumulated knowledge: 'User prefers concise responses'
  • Domain facts extracted from documentation

Counterexamples

Things that might seem like Semantic Memory but are not:

  • Remembering 'On Tuesday, user asked about React' (that's episodic)
  • The model's training knowledge (that's parametric, not semantic memory)
  • Procedural knowledge like how to use a tool

Relations

See Also

Implementations

Tools and frameworks that implement this concept: