Deterministic Memory Fusion for Long-Horizon Conversational Agents

Abstract

Long-horizon conversational agents require memory fusion to consolidate redundant information, but current approaches rely on LLM-guided fusion that introduces API costs, latency, and non-deterministic outputs. We propose DFM-Fusion, a deterministic four-stage pipeline that replaces LLM-guided fusion with sentence segmentation, near-duplicate removal via embedding similarity, MMR-based sentence packing, and salient-token coverage verification. On the LoCoMo benchmark, DFM-Fusion achieves 106.4% gap recovery on multi-hop F1 (18.72 vs 18.63, $p=0.864$), demonstrating statistical equivalence to LLM-guided fusion while eliminating all 226 fusion-related LLM calls per run. The approach provides 5.23$\times$ speedup in memory maintenance operations and guarantees verbatim quote preservation through fully deterministic, auditable fusion.