WASP-Mem: Let Agent Organize Procedural Memory. Human Spend Effort on Diagnosis

| Source: arXiv AI

Tags: WASP-Mem, Procedural Memory, TheAgentCompany, LLM Agents, Enterprise Automation

WASP-Mem has agents distill execution traces into linked wiki pages as procedural memory, raising TheAgentCompany checkpoint pass rates by 4.6–6.1% over the best baseline with 57–61% less storage.

Details

LLM agents struggle with enterprise procedures because pre-built harnesses miss proprietary policies, undocumented behavior, and state-dependent recovery paths that only show up during execution. WASP-Mem (Wiki-based Agent-Synthesized Procedural Memory) is training-free: it distills execution traces and feedback into persistent, interlinked wiki pages that consolidate successes and diagnosed failures into reusable actions and recovery strategies. On TheAgentCompany benchmark it raised checkpoint pass rate by 4.6–6.1% over the strongest baseline across three models while using 57–61% less storage. A controlled study covers consolidation method, feedback type, trace scope, and refinement rounds. The authors report that human-defined schemas and explicit deduplication add less than expected, suggesting people should spend effort on diagnosis while the agent organizes memory.