Designing memory for agents

Agents that talk to buyers and suppliers keep meeting the same people and the same problems, but start every conversation from zero. Stuffing everything into the prompt doesn't scale, and a vector search over everything returns plausible but stale context.

This was design and research work, and it has not shipped. We specified the architecture and data structure for a CRM that doubles as a unified memory backend for agents. The design separates three kinds of memory: episodic (what happened), semantic (what is true) and procedural (how to do something). Retrieval uses progressive disclosure, so an agent sees short descriptions first and pulls full detail only when it needs it. Older memories fade with time decay, and keyword-style retrieval was studied as a simpler, more predictable alternative to vector-based RAG.

We also explored AI-generated trigger tests: generate the requests that should and shouldn't pull in a given skill or memory, then tune its description until it fires on the right ones. And we reviewed the evaluation method and experiment design for this work, so the results could be trusted when it moves forward.

A specification and a set of design decisions the team can build on, not a shipped system. Pick a query and move the decay slider to see which memories would surface: episodic ones drop out first, procedural ones never fade.