Causal Memory Selection Improves Long-Horizon LLM Agents
A new paper on arXiv proposes Causal Memory Intervention (CMI), a technique for long-horizon LLM agents that selects memories based on their causal effect on task performance, rather than relying on semantic similarity. The authors introduce Causal-LoCoMo, a benchmark with causally annotated conversational data including useful memories, irrelevant distractors, and synthetic harmful memories. CMI is compared against vector and graph-based retrieval methods.
Key facts
- CMI estimates how candidate memories affect model answers under controlled interventions.
- Causal-LoCoMo benchmark is derived from long conversational data.
- Memories can be topically related but irrelevant, stale, or misleading.
- Existing memory systems treat retrieved memories as uniformly useful.
- CMI suppresses unstable, irrelevant, or harmful memories.
- The paper is available on arXiv with ID 2605.17641.
- CMI is compared against vector and graph-based methods.
- The approach targets long-horizon LLM agents with persistent memory.
Entities
Institutions
- arXiv