Semantic Memory
The AI agent memory tier that stores facts and preferences extracted from many interactions — 'this customer uses Salesforce, prefers Slack over email, is in the enterprise tier.' Stored as structured attributes (or in a knowledge graph like Zep uses), updated when new information contradicts old, and typically injected into the system prompt at conversation start. Semantic memory is what makes an agent feel like it *knows* the user rather than reading their whole history every turn. Distinct from Episodic Memory (specific events) and Working Memory (current context).
Frequently asked questions
- What happens when two extracted facts contradict?
- This is the reconciliation problem. Simplest strategy: recency wins — new facts overwrite old. Slightly better: keep a small audit history and score facts by confidence and source. Frameworks like Mem0 and Zep handle this automatically; roll-your-own implementations often skip reconciliation and quickly develop inconsistent user profiles. If you're rolling your own, either pick a rule and enforce it, or reach for a framework.