TEPA: Revoking Stale Memories for Conflict-Robust Language Agents
TEPA treats memory validity as a state agents can revoke, not just overwrite.
The paper defines “memory pollution” as stale stored evidence continuing to surface after newer conflicting evidence arrives. TEPA stores observations as keyed precedents, revokes contradicted active entries under the same key, and keeps revoked history for audit. In reversal tests, TEPA stayed at 0.950 while append-only and last-write-wins memory fell to 0.210, below no memory. The authors say its limits show up in multi-hop and very long-context settings, where retrieval chains and context selection become the bottleneck. ArXiv · AI/CL/LG's note
The paper defines “memory pollution” as stale stored evidence continuing to surface after newer conflicting evidence arrives. TEPA stores observations as keyed precedents, revokes contradicted active entries under the same key, and keeps revoked history for audit. In reversal tests, TEPA stayed at 0.950 while append-only and last-write-wins memory fell to 0.210, below no memory. The authors say its limits show up in multi-hop and very long-context settings, where retrieval chains and context selection become the bottleneck. ArXiv · AI/CL/LG's note
score 4