Last-Write-Wins Memory: Isolating Deterministic Overwrite Semantics for Long-Context Conflict Resolution

Abstract

LLM agents require long-term memory to maintain knowledge across extended interactions, yet real-world facts change over time, creating conflicting versions in memory stores. Existing systems either preserve all versions (append-only) or rely on implicit recency signals, both of which fail for multi-hop reasoning where stale intermediate facts corrupt reasoning chains. We propose Last-Write-Wins Knowledge Objects (LWW-KO), a memory system that applies deterministic overwrite semantics---filtering stale fact versions before retrieval---to resolve conflicts at the source. Through a controlled three-condition experiment on the FactConsolidation benchmark at 262K tokens, we isolate the effect of overwrite semantics from structured extraction. LWW-KO improves multi-hop accuracy by 13 percentage points over append-only memory ($p = 0.0003$) while maintaining single-hop performance. Error analysis reveals that 33% of baseline errors are stale-answer errors, and LWW-KO's resolution of these accounts for 73% of its improvement. Our results exceed all published baselines by +15pp on multi-hop and +18pp on single-hop.