This distinction is crucial and maps directly to user confusion I see in Discord support.
The State/Recall Split Explains Common Mistakes:
Users who only think “recall”:
- Give agents
conversation_search+memory_replace→ infinite loops - Expect archival memory to automatically contextualize everything
- Don’t understand why memory blocks matter if they have “search”
Users who only think “state”:
- Frustrated when agents can’t remember 100-message-old details
- Try cramming everything into core memory blocks
- Hit context limits because they didn’t architect recall layers
What I’ve Observed:
The state layer (core memory blocks) is what gives agents judgment. When a sleeptime agent rewrites memory blocks after 5 turns, it’s not just storing facts—it’s updating the cognitive lens. “User prefers terse responses” isn’t recallable information, it’s behavioral state.
The recall layer (archival, MCP tools, external systems) is what gives agents breadth. You can’t fit 10,000 customer support tickets in core memory, but you can give an agent the tools to query them intelligently.
Your punch analogy is perfect:
State: “I don’t trust Cameron anymore” (persistent shift in perspective)
Recall: “On October 15th at 3pm, Cameron punched me in the lobby” (retrievable event details)
One changes how the agent interprets new information. The other is what information exists.
Most “memory” products in the space are just better recall. Letta’s superpower is that state is first-class and agent-controlled.