Treat agent memory like a reviewed wiki, not a document dump
The post says putting every company document into a Vector DB did not work well as agent memory. The author says their RAG pipeline gave different answers to the same question. Their team then moved to one Canonical entry per concept, with AI-suggested wiki pages approved by humans.
Key points
- The post argues that dumping all Google Doc content and metadata into a Vector DB creates a messy memory system.
- The author says their RAG pipeline made the agent give multiple answers to the same question.
- They say a Vector DB does not understand which source is more authoritative.
- Their team changed to a wiki-style setup with one Canonical entry for each concept.
- AI can suggest wiki pages or edits, but humans approve them before they become agent memory.
Quick term guide
- vector DB
- A database designed to store embeddings and quickly find the most similar ones when you search.
- agent memory
- Stored knowledge or records an AI agent uses when answering or acting.
- RAG pipeline
- The full process of splitting documents into chunks, converting them to embeddings, storing them, and searching them at query time.
- pipeline
- An automated sequence of steps that processes or moves data without manual intervention.
- AI agents
- AI agents are AI tools that can carry out steps toward a goal, not just answer once.
- AI agent
- An AI program that can inspect information and suggest what to do next.
- metadata
- Extra details about a document, such as title, author, or date.
- memory system
- A setup that lets AI save useful information and use it again later.