文章背景与核心概要
长期记忆系统使语言智能体能够复用过去的的事实、偏好和任务经验。然而，持久化存储也带来了一个显著的可证伪性挑战：当外部世界发生变化时，陈旧的记忆往往仍然能够被检索到，并污染智能体的提示词（Prompt）。
本文将这种失效模式定义为记忆污染（memory pollution）——即被更新、冲突的证据所取代的活跃记忆所导致的性能退化。为了解决这一问题，作者引入了 TEPA（一种可撤销的证据-记忆机制），将有效性视为记忆的显式状态： * 键控先例（Keyed Precedents）： 观察结果被表示为与特定键绑定的结构化先例。 * 生命周期撤销（Lifecycle Revocation）： 当新鲜证据与同一键下的现有先例产生冲突时，TEPA 会自动撤销该活跃先例。 * 可审计性（Auditability）： 尽管检索严格限于当前证据，但被撤销的历史记录仍保持完整，以便进行审计和未来的重新提升（re-promotion）。
TEPA: Revoking Stale Memories for Conflict-Robust Language Agents
Authors: Yan Zhou, Yue Ouyang, Kaiyang Zheng, Suncheng Xiang
Published: August 7, 2026
Primary Subject: Artificial Intelligence (cs.AI)
arXiv: 2608.07429 | DOI: 10.48550/arXiv.2608.07429
Executive Summary
Long-term memory systems allow language agents to reuse past facts, preferences, and task experiences. However, persistence introduces a significant falsifiability challenge: when the surrounding world changes, stale memories often remain retrievable and pollute the agent's prompt.
This paper characterizes this failure mode as memory pollution—degradation caused by active memories that have been superseded by newer, conflicting evidence. To combat this, the authors introduce TEPA, a revocable evidence-memory mechanism that treats validity as an explicit state of memory: * Keyed Precedents: Observations are represented as structured precedents tied to specific keys. * Lifecycle Revocation: When fresh evidence contradicts an existing precedent under the same key, TEPA automatically revokes the active precedent. * Auditability: While retrieval is restricted exclusively to current evidence, revoked history remains intact for auditing and future re-promotion.
核心结果与性能表现
- 受控隐式机制漂移（Controlled Hidden-Regime Drift，50个随机种子）： 在完全逆转的条件下，传统的追加写入（append-only）和最后写入胜出（last-write-wins）记忆系统的性能严重退化（两者的得分均为
0.210，低于无记忆基线的0.309）。TEPA 以0.950的高分显著优于它们。 - 真实文件支持的可执行漂移（Real File-Backed Executable Drift）： 在真实文件执行条件下复制了完全相同的性能模式（追加写入：
0.203，无记忆：0.298，TEPA：0.950）。 - MemoryAgentBench (SH-6k)： 在干净的单跳（single-hop）任务上，TEPA 的表现与强大的最后写入胜出缓存相匹配，这验证了当前键替换（current-key replacement）可以作为单跳事实整合的最佳决定性操作。
- 边界局限性（Boundary Limitations）： 对多跳（multi-hop）和长上下文 MemoryAgentBench 配置的边界测试突显了检索链和上下文选择的瓶颈，这些瓶颈超出了简单的单事实有效性跟踪的范畴。
Key Results & Performance
- Controlled Hidden-Regime Drift (50 seeds): Under full reversal conditions, traditional append-only and last-write-wins memory systems degraded severely (both scoring
0.210, falling below a baseline with no memory at0.309). TEPA significantly outperformed them with a score of0.950. - Real File-Backed Executable Drift: Replicated the exact same performance patterns under real file execution conditions (append-only:
0.203, no memory:0.298, TEPA:0.950). - MemoryAgentBench (SH-6k): On clean, single-hop tasks, TEPA matches a strong last-write-wins cache, validating that current-key replacement functions as the optimal decisive operation for single-hop fact consolidation.
- Boundary Limitations: Boundary tests on multi-hop and long-context MemoryAgentBench configurations highlight retrieval-chain and context-selection bottlenecks that go beyond simple fact-level validity tracking.
结论
TEPA 将生命周期撤销确立为自主智能体的一项至关重要的核心记忆操作。通过显式管理存储知识的有效性状态，智能体能够有效地证伪过时数据、维护可靠的审计跟踪，并准确地适应不断演进的环境。
Conclusion
TEPA establishes lifecycle revocation as an essential core memory operation for autonomous agents. By explicitly managing the validity states of stored knowledge, agents can effectively falsify outdated data, maintain reliable audit trails, and accurately adapt to evolving environments.
