Vector Memory Store

An agent's context window is small and the things it ought to remember are not. The usual way out is to keep memories outside the model and fetch only the relevant ones back. A vector memory store is the database that makes that possible: text is turned into embeddings — numeric vectors that place similar meanings near each other — and stored so the agent can later retrieve by similarity rather than by exact keyword.

How the agent uses it

• Write — a conversation turn, tool result or extracted fact is embedded and stored with metadata: who, when, which session, which task.
• Read — the current step is embedded too, and the nearest stored vectors come back as candidate memories.
• Inject — the top results are placed into the prompt, so the model sees a handful of relevant memories instead of an entire history.

Why it matters, and where it bites
This is what lets an agent recall a preference stated three sessions ago, or avoid repeating a tool call that already failed. Chroma, Pinecone, Qdrant, FAISS and pgvector all serve this role, usually behind a framework's memory interface. The failure modes are specific: similarity is not relevance, so a semantically close but outdated memory is retrieved with full confidence; nothing expires unless something deletes it, so contradictory facts pile up; and pulling back too many chunks crowds the very context window the store was meant to protect. A working memory layer needs write discipline — what is worth storing, scoped how, superseded by what — as much as it needs the index.