Harrison Chase identifies three types of long-term memory (mapping from human psychology):
- Semantic memory — facts in a database, searched via semantic search. “We know how to do that.”
- Episodic memory — previous interactions. “That’s just like giving the agent a tool to look at its previous conversations.”
- Procedural memory — “what the agent should actually do, like its core instructions.” This includes system prompts, skills, and even tool configuration.
His ranking: “That type of procedural memory, I think that’s the highest impact. That’s the type that we’ve probably built to the most into the framework.”
Why procedural memory matters most: agents do repeated tasks. “I have an email agent, it looks at every single email. If I correct it once, I don’t want to correct it a second time or third time.” Semantic memory retrieves facts. Episodic memory recalls what happened. Procedural memory determines what the agent does differently next time.
The update mechanism is still open: does procedural memory update inline (agent edits its own instructions while running) or via sleep-time compute (cron job reviews traces overnight)? Chase: “We think it’s still really early, but that’s [procedural memory] where we’ve built the most.”
Connects to Agents learn at three distinct layers — model weights, harness code, and context configuration — procedural memory maps to the context layer. Also connects to Compound engineering makes each unit of work improve all future work — procedural memory is the mechanism by which each correction compounds.