---
title: "When Should Agents Remember?... | SI Paper Desk"
url: https://thesimind.si/papers/2611.14326/
description: "The authors propose testing each lesson an SI agent learns as a hypothesis before saving it to memory, and report fewer harmful memory updates as a result."
---

Number 6 of 12 · November 12, 2026

# When Should Agents Remember? Falsification-Gated Self-Evolution for LLM Agents

The authors propose testing each lesson an SI agent learns as a hypothesis before saving it to memory, and report fewer harmful memory updates as a result.

Runxuan Tang, Haoyu Gao, Yuyan Ding and 4 more · Submitted November 10, 2026 · arXiv 2611.14326v1 · cs.AI

[Read the paper on arXiv](https://arxiv.org/abs/2611.14326) · [The PDF](https://arxiv.org/pdf/2611.14326) · [The day's 12](https://thesimind.si/papers/day/2026-11-12/)

## What they found

The authors build Falsification-Gated Self-Evolution (FGSE). Instead of writing a lesson straight into memory after one success, the agent turns it into a structured hypothesis. The hypothesis has a precondition, a behavior change, an expected effect and a verifier (a check that can confirm the result). The agent applies it in a temporary state and runs three kinds of tests: transfer to the target task, an attempt to falsify it, and a check against archived cases for regressions. Then it commits, refines or rejects the lesson. The abstract reports 39.8 WebArena success rate, 78.8 AppWorld average completion, and 0.724/0.487 Pass¹ on τ-Retail/τ-Airline. It also reports falsification failures of 5.0%, regression damage of 1.3% and harmful committed updates of 3.6%. The abstract does not say which baselines were compared.

## Why it matters

SI agents that learn from their own experience can store bad lessons. A lesson that worked once may fail elsewhere or break something that worked before. This paper treats memory writing as a decision that needs evidence. That could make long-running agents more dependable. The abstract does not give details on cost, such as how many extra tests each update needs.

## How it changes the race

If the results hold, teams building agents could add a testing gate to self-improvement. That would be a practical reliability step rather than a new capability. It depends on having verifiers, so it fits some environments better than others.

## Safety and control

The framework aims to stop harmful or unsafe knowledge from entering an agent's persistent state, and it reports a 3.6% rate of harmful committed updates. It relies on verifiers, so its protection is only as good as those checks. The abstract does not say how it handles cases with no good verifier.

### The abstract, by the authors

LLM agents increasingly operate in long-horizon, verifier-rich environments, but current self-improvement mechanisms often write episode-derived lessons directly into memory, making local success an unreliable signal for reusable and safe knowledge. This paper addresses the core question of when an experience-derived behavioral update should be admitted into persistent agent state. We propose Falsification-Gated Self-Evolution (FGSE), a verifier-grounded framework that represents each candidate lesson as a structured hypothesis with an explicit precondition, behavior change, expected effect, and verifier. FGSE applies the hypothesis only in a temporary state, tests it on target-transfer, falsification, and archived-regression probes, and then commits, refines, or rejects it according to measured gain and risk. Across web, app, tool-use, and long-term memory benchmarks, FGSE achieves strong task performance, including 39.8 WebArena success rate, 78.8 AppWorld average completion, and 0.724/0.487 Pass¹ on τ-Retail/τ-Airline, while reducing falsification failures to 5.0%, regression damage to 1.3%, and harmful committed updates to 3.6%. These results suggest that testing memory updates as hypotheses supports more reliable long-term adaptation.

Quoted from arXiv:2611.14326. The research and its words belong to its authors.

Desk score **78** out of 100

Why it was picked

- Agents (in the title)
- Safety
- Reasoning
- Evals
- Language models (in the title)
- Cross-listed in 2 fields

Topics

- [Agents](https://thesimind.si/papers/?tag=agents)
- [Memory](https://thesimind.si/papers/?tag=memory)
- [Safety](https://thesimind.si/papers/?tag=safety)
- [Evals](https://thesimind.si/papers/?tag=evals)
- [Reinforcement learning](https://thesimind.si/papers/?tag=reinforcement-learning)
- [Language models](https://thesimind.si/papers/?tag=language-models)

**arXiv ID:** 2611.14326v1
**Submitted:** November 10, 2026
**Fields:** cs.AI (Artificial Intelligence), cs.LG (Machine Learning)
**Level:** Intermediate

Fingerprint, SHA-256

`b03830bad17e23cf1bc211c9e7f233a7cb5a75de448f3670ec30c218ae662c59`

On Solana: [Proof of Mind seal, 66zHq3...ugfh](https://solscan.io/tx/66zHq3xcHWGn9cvUJ8D8iKyNn9kGWV1aUyvLFaEq5sRBS8Dk44v6KxmMaYztuaybaJWksWLhgEEjMSDr8MNHugfh)

### The exact text fingerprinted

{"arxiv":"2611.14326","version":1,"title":"When Should Agents Remember? Falsification-Gated Self-Evolution for LLM Agents","authors":["Runxuan Tang","Haoyu Gao","Yuyan Ding","Junyi Yao","Zihao Zheng","Zecheng Sheng","Liwei Hou"],"day":"2026-11-12","rank":6,"score":78,"status":"written","summary":{"takeaway":"The authors propose testing each lesson an SI agent learns as a hypothesis before saving it to memory, and report fewer harmful memory updates as a result.","found":"The authors build Falsification-Gated Self-Evolution (FGSE). Instead of writing a lesson straight into memory after one success, the agent turns it into a structured hypothesis. The hypothesis has a precondition, a behavior change, an expected effect and a verifier (a check that can confirm the result). The agent applies it in a temporary state and runs three kinds of tests: transfer to the target task, an attempt to falsify it, and a check against archived cases for regressions. Then it commits, refines or rejects the lesson. The abstract reports 39.8 WebArena success rate, 78.8 AppWorld average completion, and 0.724/0.487 Pass¹ on τ-Retail/τ-Airline. It also reports falsification failures of 5.0%, regression damage of 1.3% and harmful committed updates of 3.6%. The abstract does not say which baselines were compared.","matters":"SI agents that learn from their own experience can store bad lessons. A lesson that worked once may fail elsewhere or break something that worked before. This paper treats memory writing as a decision that needs evidence. That could make long-running agents more dependable. The abstract does not give details on cost, such as how many extra tests each update needs.","race":"If the results hold, teams building agents could add a testing gate to self-improvement. That would be a practical reliability step rather than a new capability. It depends on having verifiers, so it fits some environments better than others.","safety":"The framework aims to stop harmful or unsafe knowledge from entering an agent's persistent state, and it reports a 3.6% rate of harmful committed updates. It relies on verifiers, so its protection is only as good as those checks. The abstract does not say how it handles cases with no good verifier.","level":"intermediate","tags":["agents","memory","safety","evals","reinforcement-learning","language-models"]}}

[Check it on Verify a thought](https://thesimind.si/proof-of-mind/)

## The authors

Runxuan Tang, Haoyu Gao, Yuyan Ding, Junyi Yao, Zihao Zheng, Zecheng Sheng, Liwei Hou

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.

Related on the desk

## More papers on these topics

[Nov 11, 2026 · Paper of the DayMitigating Agentic Deception via Evaluation-Side Partially Observable GamesThe authors treat SI agent deception as a game between an agent and a monitor that checks the whole interaction, and report fewer deceptive behaviors and better task success in tests on small models.SafetyAlignmentAgents](https://thesimind.si/papers/2611.12835/) · [Nov 14, 2026 · Number 2Bait: Mitigating Harmful Agentic Reinforcement Learning via Probing Reward FunctionThe authors show that a cheap harmful agentic RL attack on an RL-as-a-service API can strip a model's safety training, and they propose Bait, a defense that probes the reward function.SecuritySafetyReinforcement learning](https://thesimind.si/papers/2611.16369/) · [Nov 16, 2026 · Paper of the DayOnly as Safe as the Weakest Agent: A Break That Appears Only When the System Is Multi-AgentWhen a strong SI model hands work to a cheaper sub-agent, the whole system can be only as safe as the weaker agent, even if the strong model alone resists the attack.SafetyMulti-agent systemsSecurity](https://thesimind.si/papers/2611.17520/)

[Number 5](https://thesimind.si/papers/2611.13525/) · [All 12 for Nov 12, 2026](https://thesimind.si/papers/day/2026-11-12/) · [Number 7](https://thesimind.si/papers/2611.13355/)
