Hindsight Memory-PRM: Supervising Memory Management with Auditable Hindsight Credit
The paper trains LLM agents to manage memory using hindsight evidence from retrievals and citations instead of human labels.
Hindsight Memory-PRM turns an agent’s audit trail into credit for memory actions, including a controlled deletion-and-reanswer probe to estimate whether a stored entry mattered. The authors report that a local 8B policy reached 77.5% on held-out LoCoMo, above its API teacher’s 65.1%, while using less context than Mem0’s cited operating point. They also report 79.0% on LongMemEval, with ablations pointing to causal calibration as the source of the gain. Source: ArXiv · AI/CL/LG's note.
Hindsight Memory-PRM turns an agent’s audit trail into credit for memory actions, including a controlled deletion-and-reanswer probe to estimate whether a stored entry mattered. The authors report that a local 8B policy reached 77.5% on held-out LoCoMo, above its API teacher’s 65.1%, while using less context than Mem0’s cited operating point. They also report 79.0% on LongMemEval, with ablations pointing to causal calibration as the source of the gain. Source: ArXiv · AI/CL/LG's note.
score 5