More Memory, Worse Agents: Error Reproduction and Anti-Persistence in LLM Agents

Ruhan Wang; Kishan Panaganti; Dongruo Zhou

cs.AIOct 19, 2026

More Memory, Worse Agents: Error Reproduction and Anti-Persistence in LLM Agents

Authors: Ruhan Wang, Kishan Panaganti, Dongruo Zhou

Organizations: Indiana University · Tencent HY

Abstract

LLM agents can improve across tasks without updating model parameters by storing experience from their own executions in external memory and reusing it to guide later decisions. A central design question is whether to retain only observed trajectories and outcomes or also the rules, summaries, and skills inferred from them. Retaining these abstractions allows agents to reuse strategies without repeatedly deriving them from raw trajectories. However, errors in stored abstractions can propagate to subsequent decisions through retrieval and reuse. They can also shape execution trajectories, which then become the basis for new abstractions that reproduce the original errors. We term this feedback mechanism error reproduction: errors are regenerated through a cycle of abstraction, execution, and re-abstraction, extending their persistence beyond direct reuse of the original faulty knowledge. We propose Anti-Persistence, a memory design that accumulates trajectories and outcomes while reconstructing task-specific guidance on demand. An adaptive selector uses prior outcomes to decide whether and how to synthesize guidance, which is discarded after execution. This prevents its direct reuse as a stored object, although indirect feedback through retained trajectories remains possible. Across different benchmarks and backbones, Anti-Persistence achieves higher overall held-out success and the lowest average LLM-call cost among the evaluated memory designs. These results support separating experience accumulation from abstraction persistence in memorybased self-improvement.

Figures & tables

Figure 1: Error propagation and reproduction in agent memory. Left: An illustrative error in a stored trajectory is reused through retrieval. Center: A faulty inferred rule may shape later trajectories and be encoded again through subsequent synthesis. Right: Possible routes of influence. Both designs can retain feedback through stored trajectories; the examples imply neither a fixed ordering of their reach nor inevitable amplification.

	Llama-4-Scout			Qwen3-Coder-Next			Gemma-4-31B-it
Method	ALF	Web	Sci	ALF	Web	Sci	ALF	Web	Sci
ReAct	29.3 $\pm$ 3.1	33.3 $\pm$ 1.2	9.4 $\pm$ 1.0	70.4 $\pm$ 1.8	36.0 $\pm$ 2.0	27.2 $\pm$ 2.6	72.7 $\pm$ 1.2	43.3 $\pm$ 2.3	44.2 $\pm$ 1.2
Reflexion	46.7 $\pm$ 6.1	45.3 $\pm$ 4.2	22.8 $\pm$ 1.0	87.3 $\pm$ 3.1	47.3 $\pm$ 1.2	33.9 $\pm$ 3.5	98.0 $\pm$ 0.8	50.0 $\pm$ 2.0	44.5 $\pm$ 2.6
ExpeL	7.2 $\pm$ 3.2	38.0 $\pm$ 5.3	5.0 $\pm$ 1.9	88.0 $\pm$ 2.0	47.3 $\pm$ 2.3	32.8 $\pm$ 4.8	96.7 $\pm$ 2.3	38.6 $\pm$ 1.9	51.3 $\pm$ 1.2
AutoRefine	36.3 $\pm$ 3.2	37.3 $\pm$ 2.3	15.3 $\pm$ 1.7	83.7 $\pm$ 6.7	40.7 $\pm$ 2.3	27.8 $\pm$ 1.0	88.0 $\pm$ 4.0	48.0 $\pm$ 4.0	43.3 $\pm$ 0.0
Traj. Mem.	41.6 $\pm$ 1.7	45.3 $\pm$ 1.2	20.0 $\pm$ 1.7	84.4 $\pm$ 2.6	50.7 $\pm$ 1.2	43.8 $\pm$ 3.7	96.0 $\pm$ 1.0	48.0 $\pm$ 0.2	51.7 $\pm$ 1.7

Table 1: Held-out task success rate (%, mean

\pm

std over 3 seeds) on ALFWorld, WebShop, and ScienceWorld with three backbones. Anti-persistence is compared with agent baselines and persistent-memory variants.

Panel of Figure 2 — Figure 2: Per-iteration performance of four memory configurations over ten memory-update iterations across three benchmarks and two backbones.

Figure 3: Mean LLM-call cost and held-out success across nine settings.

Model	Bench	Discard (Ours)	Persist	$\Delta$
Llama-4-Scout	ALF	54.7 $\pm$ 2.9	26.7 $\pm$ 6.1	+28.0
Llama-4-Scout	Web	46.3 $\pm$ 1.8	34.0 $\pm$ 4.0	+12.3
Qwen3-Coder-Next	ALF	96.0 $\pm$ 1.4	92.7 $\pm$ 1.2	0 +3.3
Qwen3-Coder-Next	Web	54.0 $\pm$ 2.0	47.3 $\pm$ 3.1	0 +6.7

Table 2: Held-out success with abstractions persisted or discarded (%, mean

\pm

s.d.; 3 seeds).

Panel of Figure 8 — Figure 8: Mode-selection frequencies averaged over iterations $1$ – $10$ . Panels denote benchmarks, rows denote backbones, and columns denote modes ( $\mathrm{none}$ , imitation, strategic, corrective). Darker cells indicate more frequent selection.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Parameter (symbol)	Default	Description
Bandit: mode selection
KNN neighborhood size $\hat{B}$	$10$	top- $\hat{B}$ entries form $\mathcal{N}_{\hat{B}}(\tau)$
UCB coefficient $c$	$0.5$	exploration bonus base
Recency decay $\gamma$	$0.998$	weight $w_{i}\propto\gamma^{\,t_{\max}-t_{i}}$
KNN trust threshold $\rho$	$0.7$	fallback to global-best if $\overline{\mathrm{sim}}<\rho$
Per-mode warmup $W_{\mathrm{init}}$	$3$	each mode forced at least $W_{\mathrm{init}}$ times before KNN-UCB

Appendix

Table 3: Method hyperparameters and default values. Sensitivity analyses vary the specified parameters.

Benchmark	Training tasks	Held-out tasks	Evaluation set
ALFWorld	$180$	$134$	$134$ unseen tasks [ 18 , 35 ]
WebShop	$200$	$100$	$100$ held-out instructions
ScienceWorld	$150$	$60$	$2$ held-out instances per task type

Appendix

Table 4: Per-benchmark training and held-out task counts used in this study.

Model	Inference hardware	Context length
Llama-4-Scout [ 10 ]	$2\times$ NVIDIA H100	64K
Qwen3-Coder-Next [ 2 ]	$4\times$ NVIDIA H100	256K
Gemma-4-31B-it [ 8 ]	$4\times$ NVIDIA L40S	256K

Appendix

Table 5: Backbone configurations.

Panel of Figure 9 — Figure 9: Per-iteration performance on Gemma-4-31B-it across the three benchmarks.

	Llama-4-Scout			Qwen3-Coder-Next			Gemma-4-31B-it
Method	ALF	Web	Sci	ALF	Web	Sci	ALF	Web	Sci
Traj. Mem.	17.87 $\pm$ 0.40	4.21 $\pm$ 0.09	5.25 $\pm$ 0.10	10.90 $\pm$ 0.15	6.12 $\pm$ 0.03	11.41 $\pm$ 0.32	11.87 $\pm$ 0.19	4.85 $\pm$ 0.06	13.00 $\pm$ 0.05
Abs. Mem.	16.74 $\pm$ 4.77	4.53 $\pm$ 0.27	1.33 $\pm$ 1.15	13.64 $\pm$ 0.03	6.91 $\pm$ 0.10	15.08 $\pm$ 1.05	13.46 $\pm$ 1.04	6.04 $\pm$ 0.17	13.80 $\pm$ 0.59
Joint Mem.	19.11 $\pm$ 0.48	3.23 $\pm$ 1.57	1.15 $\pm$ 0.86	10.95 $\pm$ 0.33	7.21 $\pm$ 0.10	13.93 $\pm$ 0.95	11.01 $\pm$ 0.50	5.96 $\pm$ 0.17	12.78 $\pm$ 0.71
Anti-pers.	15.13 $\pm$ 0.19	3.99 $\pm$ 0.05	3.56 $\pm$ 0.22	11.84 $\pm$ 0.28	6.64 $\pm$ 0.55	12.68 $\pm$ 0.40	11.20 $\pm$ 0.38	5.30 $\pm$ 0.19	13.41 $\pm$ 0.19

Appendix

Table 6: Average LLM-call cost per task after the final memory-update iteration (mean

\pm

standard deviation over

3

seeds). Lower is better.

Baseline	Mean $\Delta$	95% CI	Paired $t$ -test		Wilcoxon signed-rank		Positive
	(pp)		Raw $p$	Holm $p$	Raw $p$	Holm $p$	cells
ReAct	$+19.2$	$[14.1,24.4]$	$<0.0001$	$0.0002$	$0.0039$	$0.0273$	$9/9$
Reflexion	$+7.0$	$[2.7,11.4]$	$0.0060$	$0.0179$	$0.0078$	$0.0273$	$8/9$
ExpeL	$+14.9$	$[3.6,26.1]$	$0.0157$	$0.0179$	$0.0039$	$0.0273$	$9/9$
AutoRefine	$+13.2$	$[8.7,17.6]$	$0.0001$	$0.0007$	$0.0039$	$0.0273$	$9/9$
Trajectory memory	$+6.4$	$[2.8,10.0]$	$0.0036$	$0.0145$	$0.0039$	$0.0273$	$9/9$

Appendix

Table 7: Exploratory comparisons across the nine fixed settings in Table 1 .

\Delta

denotes anti-persistence minus the named baseline, measured in percentage points. CIs are nominal, unadjusted 95%

t

-intervals, and all tests are two-sided. Holm correction covers the seven baseline comparisons within each test family. Shared backbones and benchmarks limit the inferential interpretation of these statistics.

Panel of Figure 12 — Figure 12: Sensitivity to trajectory retrieval budget $B$ and outcome-log retrieval budget $\hat{B}$ on Qwen3-Coder-Next. Bars show means over $3$ seeds; error bars show standard deviations across seeds.

Memory-Augmented Language Model Agents

Memory-augmented language model agents use persistent facts, interaction histories, and skills to support personalization and long-horizon tasks.

See all 444 papers

AdMem: Advanced Memory for Task-solving Agents
Jun 5, 2026Runzhe Wang, Huilin Lu, Shengjie Liu +2Continual Learning for LLM Agents LLM Agent Memory
Princeton University · Amazon · Arm
When Continual Learning Moves to Memory: A Study of Experience Reuse in LLM Agents
Apr 29, 2026Qisheng Hu, Quanyu Long, Wenya WangContinual Learning for LLM Agents LLM Agent Memory
Nanyang Technological University
Human-Inspired Memory Architecture for LLM Agents
May 8, 2026Doga Kerestecioglu, Alexei Robsky, Clemens Vasters +2Agent Memory Management LLM Agent Memory
Microsoft

Jun 5, 2026J/K move · Enter open · S save

AdMem: Advanced Memory for Task-solving Agents

Runzhe Wang, Huilin Lu, Shengjie Liu, Li Dong+1

Princeton University · Amazon · Arm

Continual Learning for LLM Agents LLM Agent Memory Persistent Memory for Language Models Procedural Memory for LLM Agents Long-Horizon Agent Tasks Multi-Agent LLMs+2

Open paper AskPDF arXiv

Large Language Models (LLMs) show promise as tool-using agents but remain limited in long-horizon tasks that require remembering, organizing, and reusing knowledge. Prior memory approaches aim to resolve the situation, but mainly focus on storing factual information. Recent work on procedural memory improves task reuse, yet often reduces to replaying past successes without addressing failure cases or online scalability. We introduce a unified and automatic memory framework that integrates semantic, episodic, and procedural memory in a bi-level design combining short-term and long-term stores. A multi-agent architecture with actor, memory, and critic agents enables automatic memory generation, reward annotation, and adaptive retrieval. Long-term memory is managed through reward-based evaluation, merging, and pruning, ensuring scalability and continual improvement. Experiments across various environments show that our approach improves robustness and success on long multi-turn tasks compared to existing baselines. This work highlights the importance of comprehensive, adaptive memory for advancing LLM-based agents.

Abstract

Memory-Augmented Language Model Agents

Explore similar work

AdMem: Advanced Memory for Task-solving Agents

When Continual Learning Moves to Memory: A Study of Experience Reuse in LLM Agents

Human-Inspired Memory Architecture for LLM Agents