Semantic Memory
Agents need 'user prefers Kanban' without replaying every chat that mentioned boards.
A wiki of facts, not a diary of every day you learned them.
Visual Workflows
Start here — scroll inside each diagram frame to explore, then use + / − to zoom up to 200% if needed.
Overview
Scroll inside the frame to explore · use + / − to zoom up to 200%
Scroll inside the frame to explore · use + / − to zoom up to 200%
Fact in, fact out
Scroll inside the frame to explore · use + / − to zoom up to 200%
Extract one claim. Embed it. Later retrieve it by meaning.
Key Takeaways
- 1.Semantic memory stores facts and meanings, not a play-by-play of events. Typical shape: embeddings in a vector store, or a fact table.
- 2.Retrieve by similarity: 'what do we know about this user / topic?'. Keep facts atomic so you can update one without rewriting a story.
- 3.Semantic memory is knowledge: names, prefs, policies. Store small claims with embeddings or keys.
- 4.Do not store a whole episode here — that is episodic.
Learn elsewhere
- →Episodic Memory
- →RAG embeddings — Phase 3
Real Example
Scenario
'Prefers Python' and 'team uses Jira' sit in semantic memory. The meeting where they said it sits in episodic memory.
What you would do
In Agent Memory, apply Semantic Memory to this scenario: 'Prefers Python' and 'team uses Jira' sit in semantic memory. Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.
Commands
Commands to Remember
Facts, not storiesAtomic claimsRetrieve by similarityUpdate one fact at a time
Cheat Sheet
Quick recap
quick ref- •Wiki, not diary
- •Embeddings or a fact table
- •Atomic updates
- •Different from episodes
Common Mistakes
- ✕Embedding entire transcripts as one 'fact'
- ✕Mixing events into the fact store
- ✕Never updating a fact when the user changes