Retrieval-Augmented Generation (RAG) is the primary memory architecture for long-horizon autonomous agents. Treating shared memory as an append-only stream introduces Semantic Shadowing, where conflicting historical observations accumulate and dominate valid recent updates.
The paper formalizes State Mutability to prove that standard dense retrieval suffers from Asymptotic Recall Decay in dynamic environments.
A Majority Vote Trap is demonstrated, showing that increasing the retrieval context window degrades generation accuracy by diluting attention under conditions of semantic equivalence.
GC-Mem Protocol
To resolve these issues, the authors introduce GC-Mem (Garbage Collection for Memory). Unlike heuristic time-decay mechanisms, GC-Mem relies on a temporal dominance operator paired with contradiction detection to surgically excise shadowed context.
Evaluation Results
The system was evaluated across a rigorous benchmark of 137,760 memory chunks and continuous accumulation sweeps. Standard RAG and timestamp re-ranking baselines experienced severe degradation.
In contrast, GC-Mem empirically recovered greater than 90% conflict resolution accuracy. The authors established strict precision and recall deployment thresholds to ensure state convergence.
Source: https://arxiv.org/abs/2609.16073
