2608.07429-TEPA: Revoking Stale Memories for Conflict-Robust Language Agents

Introduction to the show: ident: Paper Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "TEPA: Revoking Stale Memories for Conflict-Robust Language Agents".

Jane: The paper was written by Yan Zhou, Yue Ouyang, Kaiyang Zheng and Suncheng Xiang from Changsha University of Science and Technology and Shanghai Jiao Tong University.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Paper summary: Tom: So we've got a really interesting paper for you today, one that digs into something every language agent user has probably hit — the model remembers something that used to be true, and that old fact poisons the answer.

Jane: Exactly. The authors call it memory pollution. The core argument is that append-only memory — just storing every observation forever — becomes actively harmful when the world changes, because the stale fact stays retrievable right next to the new one.

Lu: And that's not just a theoretical worry. They build controlled experiments where a hidden regime flips, say the accepted tool for a task changes, and a simple append-only memory system drops below the performance of having no memory at all. That's the striking result.

Meng: The numbers are stark. In the controlled drift benchmark, append-only and last-write-wins both land at 0 point 210 success during full reversal, while a no-memory baseline sits at 0 point 309. The memory is literally making things worse than forgetting everything.

Tom: Right, and their fix is called TEPA. Instead of treating memory as just a growing pile of evidence, every observation becomes a keyed precedent with an explicit lifecycle state — active, hypothesis, revoked.

Jane: When fresh evidence contradicts an active precedent under the same key, TEPA revokes the stale one. It stays in an archive for audit, but it no longer gets retrieved into the prompt. That takes TEPA to 0 point 950 in the same reversal phase.

Lalam: What I find exciting is the reframing. The memory community has spent a lot of energy on making memories more relevant and more retrievable. This paper says relevance isn't enough — validity is a separate state you have to track. A fact can be perfectly relevant and completely wrong.

Lu: And they show the same pattern across real file-backed tool execution, preference updates, and the MemoryAgentBench benchmark. On the clean single-hop facts, TEPA matches a strong last-write-wins cache at 0 point 890.

Jane: So the question becomes — when does memory help, and when does it hurt? That's what we're going to unpack page by page, starting with how they name the disease.

Page 1 of the paper: Tom: We just set the scene, so let's look at how the paper frames the problem on the first page. The authors connect persistence to a falsifiability problem — once a memory gets retrieved, it becomes prompt evidence, and stale evidence can dominate the whole interaction.

Jane: The running example is charmingly simple. The capital of France is Paris, then later the capital of France is Lyon. Append-only memory holds both as active facts under the same key — capital of France — so at retrieval time the prompt gets a mix: Paris and Lyon.

Meng: And the model is left to sort out which one is current. The paper's point is that the memory system itself should handle that, not hope the language model notices the conflict and picks the fresh one.

Lu: That's the conflict-key check in their first figure. Two pieces of evidence share a key but assert incompatible values, so TEPA marks the older one as stale, moves it to a revoked archive, and retrieval then returns a clean context with just Lyon.

Tom: I like that the figure shows the contrast explicitly. Append-only gives you a prompt where the answer may be inconsistent. TEPA gives you a clean prompt and a consistent answer.

Jane: And the authors position this against existing memory systems — Reflexion, MemoryBank, Generative Agents, ExpeL. Those systems are great at accumulating experience, but deactivation is left implicit. Nobody really owns the operation of retiring a memory.

Lalam: That's the gap. The paper isn't claiming these systems are broken across the board. It's saying there's a missing primitive in the memory toolbox, and they're proposing to make it first-class.

Lu: Worth noting that they define memory pollution precisely — degradation caused by active memories that newer conflicting evidence has superseded. It's precise enough that they can measure it later with that pollution index.

Meng: And it connects to concept drift from the machine learning literature, but with a twist. Here the stale object is textual evidence inserted into a prompt, which makes it an operational problem at the memory layer.

Jane: So page one is all about naming the disease. The next page builds the formal machinery to diagnose it — an episode stream, hidden regimes, and a proper definition of conflict.

Page 2 of the paper: Jane: So page one named the disease, and page two gets formal. The agent interacts with an episode stream where each episode has a visible context, a post-action evidence item, and a hidden regime — that hidden regime is the ground truth the evaluator knows but the agent never sees.

Tom: The hidden regime is the key move for testing. It lets the authors define phases — stable, light drift, full reversal, partial return — and then ask how each memory method behaves as the regime flips through those phases.

Lu: Conflict is defined precisely: two evidence items conflict when they share a key but assert incompatible values. So the capital of France example becomes a formal predicate — same key, different value.

Meng: And then they define success rate per phase, plus the phase-wise memory pollution index — the gap between no memory and the method, divided by no memory. A positive value means your memory is actively hurting you in that phase.

Tom: That baseline is clever. Most memory papers compare memory systems to each other. Comparing against no memory asks a more fundamental question — is persistent storage actually earning its keep?

Jane: Right, and the failure case they're hunting is exactly the method with memory scoring worse than the method with nothing. That's the signature of pollution.

Lalam: What I appreciate is the discipline of the setup. The hidden regime stays evaluator-side. Methods never see the phase label, never see future boundaries. So when memory hurts, it's because the memory mechanism itself is broken, not because the benchmark leaked information.

Lu: The related work on this page also gets specific about benchmarks — LoCoMo, LongMemEval, MemoryAgentBench — and how they increasingly test conflict resolution, but mostly as an evaluation property rather than as a memory operation to build.

Meng: So the page does two jobs. It gives the formal language for the problem, and it situates that language against existing benchmarks where stale conflicts are becoming a recognized failure mode.

Jane: Which sets up the method nicely. We have the disease named and measured — now page three introduces the treatment, with precedents and lifecycle states.

Page 3 of the paper: Tom: Page three is the heart of the mechanism. Memory is represented as a set of precedents, and each precedent carries a key, a value, support and conflict counts, a lifecycle state, and a creation time.

Jane: The states are what's new. Hypothesis, Active, Revoked. Retrieval only ever sees active precedents. Revoked ones stay in the archive for audit but are excluded from the prompt.

Lu: The update rule uses a Beta-Bernoulli posterior — a transparent estimate of how valid a precedent is. Each piece of same-key evidence either supports the precedent or counts as a conflict against it.

Meng: And when does revocation fire? Either the posterior mean drops below a threshold after enough observations, or, once there are a few same-key outcomes, the recent success rate falls below a cutoff. The specific numbers are five total observations, three recent ones, and a recent-success cutoff of 0 point 34.

Tom: That makes intuitive sense. A single contradiction isn't enough — that could be noise. But repeated same-key contradictions push the posterior down, and eventually the precedent gets moved to Revoked.

Jane: There's also a second variant, TEPA-Full, which runs trial validation before promoting a candidate. A held-out task set checks whether injecting the candidate improves reward on support tasks, avoids harm on counterfactual ones, and doesn't contaminate unrelated domains.

Lu: That trial variant matters for preference updates, where an observed value might look good but actually be wrong. Validation is a gate before promotion, separate from revocation which cleans up what's already active.

Meng: The design contrast at the end of the page is nice — append-only keeps everything active, sliding windows cap by age, reactive forgetting clears globally after a detected failure. TEPA makes the unit of update the keyed precedent, so one revocation doesn't nuke unrelated memories.

Lalam: It's the difference between a sledgehammer and a scalpel. Reactive forgetting forgets everything when something breaks. TEPA removes only the contradicted same-key precedent and leaves everything else intact.

Meng: And because revocation is local, the audit story works too. The revoked precedent sits in the archive with its contradiction history, so you can explain why it was retired and even promote it again if the world changes back.

Jane: That re-promotion path is actually part of the lifecycle trace in the supplementary material — a precedent revoked during reversal can later be restored when evidence supports it again. That completes the loop.

Lu: Which is exactly the surgical property the experiments exploit, especially in the preference stream where the stale profile stays semantically on-topic. But before the results, page four proves why revocation should work and lays out the full experimental plan.

Page 4 of the paper: Lu: Page four has two big blocks. First, the theory — a risk argument that explains why revocation is the right operation. Under a single-key supersession model, if stale evidence gets exposed with some probability and the downstream policy follows it with some probability, then append-only memory pays an excess risk proportional to that product.

Meng: They make the condition concrete. When that stale-exposure risk term dominates the benefit of any still-current evidence, append-only memory has higher expected error than no memory at all. Revocation removes exactly the harmful term.

Tom: That's the formal backbone for the pollution index. It turns "memory can be worse than nothing" from a surprising empirical finding into a predicted outcome of stale exposure.

Jane: Then the page switches to the experimental plan — four research questions. Can append-only memory fall below no memory under hidden reversal? Does revocation prevent that? Do trials improve preference updates? And does the mechanism transfer to external benchmarks?

Lu: The benchmarks mirror those questions — controlled hidden-regime drift, real file-backed executable drift, a preference-update stream, and MemoryAgentBench SH-6k for external validation.

Meng: The baselines are thoughtful because they isolate mechanisms. No memory, append-only, temporal recency, semantic retrieval, sliding window, reactive forgetting, last-write-wins, conflict-aware recency, and oracle reset as a diagnostic upper bound that knows the phase boundaries.

Tom: That oracle reset tells you what's possible with perfect timing knowledge, and it's a smart sanity check. If TEPA beat oracle reset, you'd wonder whether something else was going on.

Jane: They also use paired statistical tests throughout, which controls for the fact that some tasks are just harder than others. Comparing methods on matched task indices isolates the memory mechanism's effect.

Lu: One detail worth flagging — the default TEPA configuration uses a proposal threshold of three, a revocation threshold of 0 point 3, and a promotion threshold of 0 point 6. And the drift benchmarks use deterministic executors, so the experiments isolate memory behavior from model sampling noise.

Meng: Which is exactly the right design for the question. They want to know whether the memory state is the culprit, and a deterministic executor makes memory state the only variable that moves.

Tom: So theory says revocation should work, and the setup is designed to catch pollution if it exists. Page five delivers the first big experimental punch — the controlled drift result.

Page 5 of the paper: Tom: And here's the result the theory predicted. In the controlled drift benchmark, during full reversal, append-only, last-write-wins, and conflict-aware recency all collapse to 0 point 210 success. No memory gets 0 point 309. That's a pollution index of 0 point 318.

Jane: Memory is actively worse than amnesia. And the paper is careful to show this is statistically solid — the paired difference between TEPA and append-only across all tasks is 0 point 167, with an effect size of 9 point 68.

Lu: TEPA sits at 0 point 950 in that same reversal phase. The authors note that errors concentrate in the first few same-key trials after reversal, before enough contradictory evidence accumulates. Once revocation fires, retrieval contains only the current precedent.

Meng: So the 0 point 950 plateau is really measuring adaptation latency. The system pays a small transition cost, then the active set is clean again. That's a nice interpretation — the metric gets tied to a concrete memory event.

Tom: And the baselines around TEPA are revealing. Reactive forgetting gets to 0 point 796, better than append-only but still well short. Oracle reset only reaches 0 point 812, because it resets at phase boundaries but still uses the same weak base executor.

Jane: That oracle number is fascinating. Perfect timing information doesn't beat TEPA's local, evidence-driven revocation. The oracle wipes memory at the wrong granularity, while TEPA surgically revokes only the contradicted key.

Lu: There's also a nice nuance in the partial return phase. When the old regime comes back, append-only recovers to 1 point 00 because its stale precedent becomes correct again, while TEPA sits slightly lower at 0 point 94 because the revoked precedent has to be re-promoted. That's the cost of lifecycle state — tiny compared to surviving the reversal.

Meng: Then they repeat the whole experiment with real file I/O — generated CSV and JSON files, real tool libraries, an accepted backend that changes for CSV tasks under reversal while JSON tasks stay stable as controls.

Tom: And the pattern reproduces almost exactly. Append-only and last-write-wins at 0 point 203, no memory at 0 point 298, pollution index 0 point 319, TEPA at 0 point 950. Reactive forgetting trails at 0 point 771.

Jane: The real-execution experiment rules out the criticism that this only happens with a symbolic reward function. The stale precedent leads the agent to call the wrong tool on actual files, and the pollution still shows up.

Lu: So both drift settings confirm the core claim. But the next setting is where things get subtler — preferences, where the stale fact is still semantically on-topic, and page six shows why they needed the trial-validated variant there.

Page 6 of the paper: Jane: The preference-update stream tests a trickier case — a user's long-term profile says one preference, fresh session feedback says the opposite. The stale profile stays semantically on-topic, so relevance-based retrieval keeps pulling it in.

Tom: And the collapse is even more dramatic. Append-only memory hits 0 point 138 during full reversal, while no memory gets 0 point 837. Last-write-wins improves to 0 point 686 overall but still falls below the no-memory baseline.

Lu: The reason last-write-wins struggles here is important. In the drift benchmarks, last-write-wins only writes successful same-key experiences, so a stale precedent survives until a new success arrives. In preference updates, the stream directly supplies asserted facts, and the stale profile stays attractive.

Meng: That's where TEPA-Full enters. The trial-validated variant reaches 0 point 910 overall and is statistically indistinguishable from no memory — a difference of 0 point 002 with a p-value of 1 point 000.

Jane: But the paper's key insight is that validation and revocation play complementary roles. Validation filters bad candidates before promotion; revocation clears already-active preferences that retrieval still finds attractive. Using one without the other leaves a gap.

Tom: Then they move to MemoryAgentBench SH-6k, the external benchmark. TEPA-Rev and last-write-wins both land at 0 point 890 on substring exact match, while removing revocation drops TEPA to 0 point 630.

Jane: Right, that ablation gap shows current-key replacement is the decisive operation when the latest fact is directly observed. The lifecycle state and archive matter most in the drift and preference settings, where valid updates have to be inferred from interaction outcomes.

Meng: And the ablation figure makes the point visually. Removing revocation collapses full-reversal success from 0 point 950 to 0 point 211, right down at append-only level, below no memory.

Lu: Hyperparameter sensitivity is low too. Threshold variants around the default stay close to zero change, and trial budgets of one, three, and five are all fine. The one negative outlier is a larger recent window, which slows adaptation.

Tom: So the mechanism is robust to its knobs. The operation itself — revocation — is what carries the weight.

Jane: Which sets up page seven's honesty check. The method hits its limits on harder tasks, and the paper is upfront about exactly where.

Page 7 of the paper: Lu: Page seven is where the paper admits boundaries, and I respect that. On the long-context single-hop setting, SH-32k, TEPA still helps at 0 point 680. But on multi-hop MH-6k it drops to 0 point 040, and on the very long SH-262k it hits 0 point 000.

Meng: Those numbers are brutal but diagnostic. Fact-level revocation repairs the validity state of a single fact before it enters the prompt. Multi-hop chain construction and very long-context selection are a different layer of the problem — retrieval planners, not memory-state tracking.

Tom: The paper frames this as a design principle — persistent memory should track validity alongside relevance. Stale evidence can be highly relevant to the current query, which is why similarity and recency give you such a weak signal for supersession.

Jane: And they're honest about the key-extraction assumption. TEPA needs useful conflict keys, a natural fit for preference slots, entity attributes, and tool-regime records. Open-ended memories with implicit conflict relations are much harder.

Lu: The key-noise audit in the supplementary material quantifies that transition. As structured key noise increases, TEPA degrades smoothly from 0 point 950 down to 0 point 777, while append-only and last-write-wins stay stuck near their reversal failure modes because they never had a lifecycle signal at all.

Meng: That's the deeper point. Even with noisy keys, TEPA retains some protection because the mechanism exists. The baselines have no mechanism to degrade gracefully from — there's no lifecycle state to corrupt.

Tom: And the conclusion restates it operationally — a memory item can be relevant, recent, and still wrong once its value has been superseded. Revocation removes it from ordinary retrieval while preserving it for audit and later re-promotion.

Lalam: The biggest picture here is recasting memory as an inspectable lifecycle. Each precedent carries a key, evidence, state, and contradiction history. An agent can explain why a memory is active, revoked, or ready to come back — that's a real shift from memory as a pile of text.

Jane: It also connects to concurrent work on stale-memory validity and provenance — STALE, Eywa, MemoryArena. This paper isolates one concrete operation while those benchmarks expand the evaluation landscape.

Lu: So the contribution is narrow but sharp — conflict-keyed revocation for stale evidence, tested across four settings with clean statistical machinery. And the boundary results tell future researchers exactly where to aim next.

Tom: Which makes it a great paper to send people to for the method, and for the honest map of what remains unsolved. Let's pull it all together for the close.

Conclusion: Tom: So let's wrap this up. The paper identified a real failure mode — memory pollution, where stale active evidence makes a memory system worse than no memory at all. The evidence across controlled drift, real file execution, and preference updates is consistent and statistically strong.

Jane: And the fix is elegant in retrospect. Make validity an explicit state of memory. Give every piece of evidence a key and a lifecycle, and when fresh evidence contradicts an active precedent under the same key, revoke it. Move it to the archive, keep it for audit, stop putting it in the prompt.

Lu: The numbers to remember: append-only at 0 point 210 versus no memory at 0 point 309 versus TEPA at 0 point 950 in controlled reversal, and the same pattern under real tool execution. On clean single-hop facts, TEPA matches the strongest replacement baseline.

Meng: The boundaries matter too. Multi-hop and very long context remain unsolved, and they point toward retrieval planners and chain construction as the next layer. What we have here is a memory-state fix, and the harder problems still need different machinery.

Lalam: I think the lasting contribution is the reframing of memory as a lifecycle with an inspectable state. That opens the door to auditing, to re-promoting old knowledge when it becomes relevant again, and to building agents that can genuinely falsify their own beliefs.

Tom: It's one of those papers where the core idea feels obvious after you hear it, but nobody had pinned it down so cleanly before.

Meng: And the measurement discipline deserves credit too — the paired tests, the pollution index, the oracle reset as a sanity ceiling. It shows how to study memory failure honestly instead of just benchmarking accuracy.

Jane: Let's leave listeners with the central image. Memory should be revocable, like evidence in a trial. It shouldn't be etched in stone. You can always bring a fact back if the world changes again.

Tom: Thanks for listening — we're done with this one, and we'll be back soon with the next paper on the arXiv.

Jane: See you then.