MemCollab: Cross-Agent Memory Collaboration via Contrastive Trajectory Distillation

arXiv·medium signal

MemCollab introduces a framework for sharing memory across different LLM agents by contrasting reasoning trajectories to distill agent-agnostic knowledge. Existing per-agent memory tightly couples stored knowledge to a single model's reasoning style, and naively transferring it degrades performance. MemCollab's contrastive process extracts shared task-level invariants while suppressing agent-specific biases, consistently improving accuracy and inference efficiency across diverse agents including cross-model-family settings on math and code benchmarks.