Structured memory vs RAG for agents
Document retrieval and agent memory solve different problems.
Retrieval-augmented generation (RAG) is excellent for pulling relevant documents into a prompt. Agent memory infrastructure is built for something else: persisting what an agent learns about a user, project, or workflow over time.
What RAG does well
RAG searches a corpus of documents — manuals, wikis, product specs — and injects matching chunks into the context window. It is stateless per query unless you add separate session logic.
What agent memory does
Archilas distills conversations into structured memory: identity, preferences, commitments, and signals. These entries update over time and surface when your agent needs them — at session start, during tool calls, or on explicit query.
Archilas architecture
- Skeleton — live working memory your agent reasons against now.
- Vault — immutable evidence log with extraction confidence scores.
- Distiller — pipeline that converts conversations into structured entries. Raw transcripts are not stored.
Use both together
RAG for your knowledge base. Archilas for what the agent remembers about each user and session. Connect via MCP or API at mcp.archilas.com and api.archilas.com.