Keep the Record, Rebuild the Rule

TL;DR for operators

An AI agent can retain a complete record of what it did without permanently storing every rule it infers from that record. This distinction matters when an instruction derived from one successful task is subsequently reused in circumstances where it no longer applies.

A new agent-memory architecture called Anti-Persistence keeps execution trajectories and outcomes but generates task-specific guidance only when needed, discarding that guidance afterward. Across nine benchmark–model combinations, it achieves 59.9% mean held-out success, compared with 53.5% for trajectory-only memory, 49.0% for abstraction-only memory, and 49.3% for joint memory.

Controlled experiments also favor discarding generated guidance over storing it. The evidence supports testing a separation between durable experience and temporary interpretation, although it does not prove that persistent instructions inevitably reproduce errors.

For enterprise agent operators, the immediate implication concerns retention policy: preserve auditable execution evidence, avoid automatically promoting generated heuristics into authoritative procedures, and require separate evidence before making instructions reusable organizational knowledge.

When a completed task becomes an operating rule

Consider an agent that repeatedly handles customer-service requests or internal administrative workflows. After completing a case, it can preserve the original request, tool observations, actions, and outcome. It can also summarize the experience into an instruction for handling similar cases.

Those records have different evidentiary status. An execution trace describes what happened in a particular situation. A generated instruction generalizes from that experience, potentially omitting conditions that made the original action appropriate.

If the instruction becomes permanent, the next agent execution may follow it even when those conditions differ. That execution can then become evidence for another generated instruction. Deleting the original rule would not automatically remove the subsequent trajectories or newly stored rules.

Ruhan Wang, Kishan Panaganti, and Dongruo Zhou examine this problem in More Memory, Worse Agents: Error Reproduction and Anti-Persistence in LLM Agents.¹ Their central design choice is to retain experience while preventing inferred instructions from automatically becoming persistent memory objects.

The concern is not memory accumulation itself. It is the possibility that an agent’s interpretation of its experience acquires the same durable status as the underlying evidence.

Anti-Persistence keeps experience available for fresh interpretation

The proposed architecture maintains task descriptions, execution trajectories, outcomes, and feedback about previous guidance choices. It does not maintain a growing library of synthesized rules.

When a new task arrives, the agent can use experience in four ways:

1. Direct reuse: Supply relevant historical records without synthesizing new guidance.
2. Imitation: Extract behavioral patterns from successful trajectories.
3. Strategic synthesis: Generate plans and subgoals from successful experience.
4. Corrective synthesis: Derive avoidance instructions from failures, optionally contrasting them with a successful execution.

The synthesized guidance can influence the current execution. Once that execution ends, the guidance is discarded, while the new trajectory and its outcome remain available.

This distinction allows the system to keep learning from accumulated evidence without treating every generated interpretation as a permanent instruction.

Choosing among the four modes requires another mechanism. The authors use a contextual-bandit controller that examines outcomes from historically similar tasks, giving greater weight to relevant and recent observations.

The controller estimates each mode’s reward after accounting for additional synthesis calls, then adds an exploration incentive for less-tested choices. Warmup, random exploration, and similarity-based fallbacks prevent it from relying exclusively on sparse local feedback.

Operationally, the agent can learn when a direct historical example is sufficient and when additional reasoning is worth its inference overhead.

Discarding generated guidance improves held-out success

The main evaluation covers ALFWorld, WebShop, and ScienceWorld, using Llama-4-Scout, Qwen3-Coder-Next, and Gemma-4-31B-it.

Each method undergoes ten memory-update iterations, with 50 sampled training episodes per iteration. Backbone parameters remain fixed. Held-out evaluations use frozen memory and controller state, and their feedback does not enter subsequent training updates. Results are averaged over three random seeds.

The final equally weighted success averages across nine benchmark–backbone settings are:

Source: Section 4.2, Table 1; Appendix F.3. These are equally weighted setting-level means, not a pooled task-level success rate.

Anti-Persistence exceeds trajectory memory by 6.4 percentage points, abstraction memory by 10.9 points, and joint memory by 10.6 points. It also leads all evaluated methods in seven of the nine settings.

Notably, abstraction-only and joint memory each trail trajectory-only memory in seven settings. Persisting additional inferred knowledge therefore does not consistently improve the agent’s ability to complete unseen tasks.

The architecture also has the lowest average LLM-call count among the four memory configurations, including execution and synthesis calls. This is an average result, not a per-setting guarantee or a complete measure of production cost.

The retention ablation isolates a consequential choice

The broader architecture comparison changes multiple components. A more targeted experiment tests whether synthesized guidance should remain available for later retrieval.

Across four model–benchmark combinations, discarding guidance improves held-out success relative to persisting it:

Source: Section 4.4, Table 2; three-seed mean differences.

These ablations give more direct support to the retention decision than the aggregate architecture comparison alone.

Separately, the adaptive controller outperforms evaluated fixed-mode and random-mode strategies on Qwen3-Coder-Next across ALFWorld and WebShop (Section 4.4, Figure 6). This supports selecting guidance according to task context, although the controller and retention policy contribute different mechanisms.

Neither result establishes universal superiority. On Gemma-4-31B-it with ScienceWorld, joint memory achieves 56.3% success against Anti-Persistence’s 54.1%.

The proposed error-reproduction mechanism is not yet causally established

The authors describe a possible feedback sequence: a stored abstraction influences an execution; the resulting trajectory is retained; later synthesis retrieves that trajectory and generates a descendant abstraction containing a related error.

An observational provenance audit finds higher operational flag rates among descendants historically exposed to flagged parent abstractions in most of six examined settings. The largest reported difference is 0.55.

That association is consistent with the proposed mechanism, but it does not isolate causal transmission. Task difficulty, retrieval selection, provenance composition, and labeling eligibility can affect the comparison. Moreover, the operational flags rely on downstream outcome associations rather than independently verified semantic errors.

The paper’s formal analysis separates three events: exposure to a flagged parent, generation of a flagged candidate, and admission of that candidate into persistent memory. Its probability decomposition shows how these events combine in one synthesis invocation. It does not establish long-run amplification or inevitable accumulation.

Anti-Persistence eliminates the final admission event for synthesized guidance. It does not prevent an incorrect temporary instruction from influencing an execution trajectory that will later be retained.

Two auxiliary experiments further qualify the interpretation. On 30 selected AppWorld tasks, useful supplied skills produce 40.0% success, versus 23.0% without skills and 20.0% with deliberately incorrect skills. Skill correctness matters for immediate execution.

On WebShop, GPT-5.5-based filtering improves a persistent abstraction-memory baseline from 38.7% to 46.0%. Yet adding a frozen filtered-skill layer to Anti-Persistence reduces its success from 56.0% to 44.0%. Filtering can help one architecture without improving another. These auxiliary results do not directly demonstrate multi-generation error reproduction.

Enterprise memory policies should distinguish evidence from authority

For teams deploying agents in recurring customer-service, administrative, or analytical workflows, Cognaptus draws a narrower operational inference than a general prohibition on persistent knowledge.

A memory system can distinguish three classes of material:

• Execution evidence: Retain task inputs, tool observations, actions, outcomes, and relevant provenance for retrieval and audit.
• Temporary interpretation: Generate guidance for the current task without automatically adding it to the organization’s reusable instruction library.
• Verified procedures: Maintain separately governed instructions whose applicability and correctness have been established through an appropriate validation process.

The third category is an architectural recommendation, not an experimentally validated feature of Anti-Persistence.

A meaningful deployment test would compare retention policies on representative, unseen organizational tasks across successive memory updates. It should measure task success, retrieval relevance, provenance, incorrect-instruction reuse, and total operating costs—not just whether an agent produces plausible lessons after completing a case.

The study provides useful evidence for such a test, but its boundary is substantial: three text-based benchmarks, three backbones, ten update iterations, and no production deployment. Its exploratory statistical comparisons also reuse benchmarks and backbones, limiting independence. LLM-call counts do not capture complete token expenditure, latency, storage, or infrastructure costs.

Persistent abstractions that have been externally verified may still offer advantages that this design deliberately forgoes.

A memory system needs a rule for promoting knowledge

The strongest finding is an architectural one: agents can benefit from interpreting past experience without permanently retaining each interpretation. In the evaluated settings, keeping execution evidence while rebuilding guidance when needed produces better average outcomes than several more persistent alternatives.

For operators, this shifts attention from how much knowledge an agent can accumulate to what evidence justifies promoting an inferred instruction into durable authority. Anti-Persistence offers one tested default for unverified guidance. Provenance, verification, indirect feedback, and production economics remain necessary parts of the broader memory-governance decision.

Cognaptus: Automate the Present, Incubate the Future.