Revoking Stale Memories for Language Agents
Listen to the summary
Uses a voice available on your device
Audio options
On this page 4 sections
Related concepts 1 concepts
Key Takeaways
- Language agents suffer from a memory pollution problem where stale active memories remain retrievable and pollute prompts as the world changes.
- TEPA introduces novel mechanics including explicit lifecycle states, Beta-Bernoulli support and conflict tracking, and archived audits for revoked precedents.
- TEPA achieves a score of 0.950 under controlled hidden-regime drift and real file execution, compared to 0.210 and 0.203 for append-only memory.
- On MemoryAgentBench SH-6k, TEPA matches a strong last-write-wins cache with a substring exact match score of 0.890.
Summary & Methodology Analysis
Language agents rely on long-term memory to reuse past facts and experiences, but persistence causes a memory pollution problem where stale active memories remain retrievable and pollute the prompt when the world changes. Standard mechanics used by prior systems include append-only memory, last-write-wins caching, temporal recency, semantic retrieval, sliding window, and reactive forgetting. TEPA addresses these limitations by introducing novel mechanics. It converts observations into keyed precedents with explicit lifecycle states, specifically Hypothesis, Active, and Revoked. It tracks support and conflict counts using a Beta-Bernoulli model to compute posterior means, revoking contradicted active precedents from ordinary retrieval while keeping them in an archive for audit, and optionally performing trial-by-execution validation on a held-out task set before promotion in a configuration called TEPA-Full.
In empirical evaluations across multiple benchmarks, TEPA demonstrates robust performance against several baselines. Under controlled hidden-regime drift across 50 seeds during full reversal, append-only and last-write-wins memory scored 0.210, no memory scored 0.309, and TEPA scored 0.950. Under real file execution, append-only memory scored 0.203, no memory scored 0.298, and TEPA scored 0.950. On MemoryAgentBench SH-6k, TEPA matches a strong last-write-wins cache with a substring exact match score of 0.890. The evaluation suite spans multiple datasets and models including MemoryAgentBench, MemoryAgentBench SH-6k, LoCoMo, LongMemEval, LongMemEval v2, MemoryBank, Generative Agents, ExpeL, SeCom, A-MEM, Mem0, MemoryOS, DeepSeek-v4-Flash, SH-32k, MH-6k, and SH-262k.
Despite its strong performance, TEPA has specific limitations noted by the authors. TEPA assumes useful conflict keys can be extracted from evidence, which is harder for open-ended memories whose conflict relation is implicit. Furthermore, boundary tests on multi-hop and very long-context settings expose retrieval-chain and context-selection bottlenecks beyond fact-level validity tracking. The paper does not specify hardware deployment requirements, dollar costs, or exact execution latency figures beyond the benchmark accuracy scores.
Interactive System Flowchart
Cross-Examination & FAQs
A deeper dive clarifying mechanics, constraints, and baseline evaluations.
Q1. What is the core problem addressed by the paper?
Language agents rely on long-term memory to reuse past facts and experiences, but persistence causes a memory pollution problem where stale active memories remain retrievable and pollute the prompt when the world changes.
Q2. What is TEPA?
TEPA is a conflict-robust memory framework that revokes stale memories and uses explicit lifecycle states and a Beta-Bernoulli model to track support and conflict counts.
Q3. How well does TEPA perform compared to standard append-only memory?
Under controlled hidden-regime drift across 50 seeds, append-only memory scored 0.210 while TEPA scored 0.950. Under real file execution, append-only memory scored 0.203 while TEPA scored 0.950.
Q4. What are the standard memory mechanics compared or utilized in the research?
The paper references standard mechanics including append-only memory, last-write-wins caching, temporal recency, semantic retrieval, sliding window, and reactive forgetting.
Q5. What are TEPA's novel mechanics?
TEPA converts observations into keyed precedents with explicit lifecycle states (Hypothesis, Active, Revoked), tracks support and conflict counts using a Beta-Bernoulli model to compute posterior means, archives revoked precedents for audit, and optionally performs trial-by-execution validation on a held-out task set before promotion in TEPA-Full.
Q6. How does TEPA perform on MemoryAgentBench SH-6k?
On MemoryAgentBench SH-6k, TEPA matches a strong last-write-wins cache with a substring exact match score of 0.890.
Q7. What models or datasets are involved in the paper?
The paper involves MemoryAgentBench, MemoryAgentBench SH-6k, LoCoMo, LongMemEval, LongMemEval v2, MemoryBank, Generative Agents, ExpeL, SeCom, A-MEM, Mem0, MemoryOS, DeepSeek-v4-Flash, SH-32k, MH-6k, and SH-262k.
Q8. What are the main limitations of TEPA?
TEPA assumes useful conflict keys can be extracted from evidence, which is harder for open-ended memories whose conflict relation is implicit, and boundary tests on multi-hop and very long-context settings expose retrieval-chain and context-selection bottlenecks beyond fact-level validity tracking.
Q9. Does the paper provide specific latency or dollar cost figures?
The paper does not specify latency or dollar cost figures.