The Immutable Past: Formalizing State Mutability and Conflict Resolution in Mutable RAG

Summary

The paper studies Semantic Shadowing in append-only RAG memory for long-horizon agents: outdated observations that are semantically equivalent to current facts accumulate and dominate both retrieval and generation. The authors formalize two theorems — Asymptotic Recall Decay (expected valid chunks in top-k tends to zero as stale history grows) and the Majority Vote Trap (attention splits across contradictory but semantically equivalent chunks, tilting output probability toward the stale majority) — and propose GC-Mem, an inference-time protocol that runs pairwise NLI contradiction detection over the retrieved candidate set and removes chunks contradicted by newer-timestamped ones (temporal dominance operator Phi_T). Experiments use a synthetic benchmark of GPT-4o-generated chat sessions (48 medium instances/1,464 chunks; 65 large/1,950 chunks) across seven LLMs, two prompt modes, and an optional short-term-memory buffer. Reported gains over standard RAG range from +4 to +46 percentage points, and a bag-of-facts ablation shows standard RAG collapsing to roughly 12% accuracy over full contradictory timelines. However, the headline Table 1 (>90% conflict-resolution accuracy) uses sample sizes (n=74/38) that match nothing in the appendix (n=65/48), with values absent from every breakdown; Table 4's timestamp-only ablation (94.7%) contradicts Table 1's timestamp baseline (10.5%); and under STM=on the paper's own aggregates show timestamp re-ranking beating GC-Mem (71.6% vs 68.1%; 67.5% vs 61.5%), undercutting the abstract. The benchmark is purely synthetic; no external-benchmark results or public code are reported.