TL;DR: Storing and retrieving raw facts works for simple Q&A, but agents that run over time need more. They need to learn from experience, adapt when things change, and connect scattered dots. That requires thinking about memory differently.

Where retrieval breaks down

We built an internal AI PM to help track projects, summarize standups, and answer questions about what's happening across teams. It has access to everything - Slack messages, meeting notes, Linear tickets, design docs, past decisions.

Ask it "what's the status of Project X?" and it retrieves relevant docs and gives you a summary. That works. But ask it "which projects are at risk?" and it struggles. The signals are there - timeline slips, engineers requesting transfers, shorter weekly updates - but no single document says "this project is at risk." The agent retrieves pieces. It doesn't connect them.

Or ask about a decision made six months ago, and it might retrieve the original proposal without noticing the follow-up thread where the team reversed course. The information is all there. The agent just doesn't know which parts are still current.

This is where I started thinking about what memory actually needs to do beyond store-and-retrieve.

There's a growing recognition of this problem. Foundation Capital recently wrote about context graphs as the next evolution beyond RAG - the idea that AI systems need to map re...

Read the full post