TEPA (Revocable Evidence-Memory Mechanism) is a system designed to solve the "memory pollution" problem in language agents, where outdated information remains active in memory and degrades the agent's performance when the environment or user preferences change. By treating memory as a set of keyed precedents with explicit lifecycle states, TEPA allows agents to revoke stale information while keeping it available for audit, ensuring that only current, valid evidence is retrieved.
The Problem: Memory Pollution
Language agents often rely on long-term memory to store facts, task experiences, and user preferences. However, when the world changes—such as a shift in tool regimes or updated user preferences—previously stored memories may become incorrect. If an agent continues to retrieve and use this stale evidence, it can perform worse than if it had no memory at all. The authors, Yan Zhou, Yue Ouyang, Kaiyang Zheng, and Suncheng Xiang, define this as "memory pollution," where the agent is actively misled by its own outdated records.
How TEPA Works
TEPA introduces a mechanism to manage the validity of memory through "keyed precedents." Each memory item is assigned a conflict key (identifying the subject) and a value. The system tracks the lifecycle of these precedents using a Beta-Bernoulli model to estimate their current validity.
When new evidence arrives that conflicts with an existing active memory under the same key, TEPA performs a state transition:
Active: The current, reliable evidence used for retrieval.
Revoked: Stale evidence that has been superseded by newer information. This is removed from the active retrieval set but preserved in an archive for auditing purposes.
Hypothesis: A state for new candidate memories before they are promoted to active status.
For more complex scenarios, the authors introduced "TEPA-Full," which uses trial-based validation. This variant tests candidate memories against a held-out set of tasks to ensure they improve performance and do not cause harm before being promoted to the active set.
Performance and Results
The authors evaluated TEPA across controlled hidden-regime drift, real file-backed executable drift, and preference-update streams. In controlled drift experiments, traditional "append-only" and "last-write-wins" memory systems performed significantly worse than having no memory at all during periods of full reversal (0.210 success rate versus 0.309 for no memory). In contrast, TEPA maintained a 0.950 success rate by effectively revoking the stale, conflicting memories. On the MemoryAgentBench SH-6k benchmark, TEPA matched the performance of strong last-write-wins caches, confirming its effectiveness in single-hop fact consolidation.
Limitations and Future Directions
While TEPA is effective at managing fact-level validity, the authors identified bottlenecks in more complex settings. Boundary tests on multi-hop reasoning and very long-context scenarios in MemoryAgentBench revealed that simply tracking the validity of individual facts is not enough. These settings expose limitations in retrieval-chain construction and context-selection, suggesting that future research must look beyond fact-level revocation to address how agents manage complex, interdependent information over long histories.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!