Zero-Mem reports structured agent memory with no LLM calls outside final question answering
- Sources: primary, discussion
- Summary: The paper describes an agent memory layer whose write, update and retrieval operations use no generation. Language model calls occur only when a final question is answered. It is a preprint with no independent replication.
- Why it matters: If memory operations need no generation, the recurring token cost of an agent's memory layer drops to encoder compute, which changes the economics of long-running sessions.