Semantic memory

Semantic memory stores abstracted, distilled knowledge: facts, heuristics, and learned conclusions. It is the "what is true" layer.

These facts come in two flavors. General facts include industry knowledge, tool capabilities, and domain terminology. User-specific facts include things like company size, tech stack, key contacts, and project deadlines. In the author's system, semantic memory is a curated MEMORY.md file in each agent's workspace.

The defining challenge of semantic memory is curation. Not everything should become a permanent fact. The "Building Long-Term Memory" article draws the line clearly: "We use Salesforce" is a durable fact worth storing, while "I'm having a busy week" is transient context that is not. The "Practical Guide" puts it bluntly: without curation, semantic memory becomes a junk drawer. This is why many systems use a staging step where candidate facts are held for review before being promoted to permanent memory, which adds a human or automated check for accuracy. Semantic memory is also where grounding comes from: instead of guessing, the agent can reference verified facts about the user's situation.