Memory as a Controlled Process: Learned Adaptive Memory Management for LLM Agents

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Eric Hanchen Jiang, Zhi Zhang, Yuchen Wu, Levina Li, Dong Liu, Xiao Liang, Rui Sun, Yubei Li, Edward Sun, Haozheng Luo, Zhaolu Kang, Aylin Caliskan, Kai-Wei Chang, Ying Nian Wu
• Venue: arXiv: cs.CL
• Published: 2026-07-15
• Collected: 2026-07-16
• URL: http://arxiv.org/abs/2607.13591v1
• PDF: https://arxiv.org/pdf/2607.13591v1
• DOI: UNKNOWN

Plain-Language Brief

The paper proposes a new framework called Memory as a Controlled Process (MemCon) for large language model (LLM) agents that use external memory systems. MemCon learns an online policy to adaptively decide when, what, and how much to retrieve from memory, when to inject distilled plans, and when to consolidate or forget based on the context of a task. The framework is backend-agnostic, meaning it can work with any existing memory implementation, and has been shown to outperform multiple memory baselines across several benchmarks.

What the Document Says

• The paper discusses MemCon, a framework for LLM agents that uses external memory systems.
• MemCon models memory operations as a Markov Decision Process (MDP).
• It learns an online policy that decides when, what, and how much to retrieve, inject plans, consolidate, or forget based on the context of a task.
• The framework is backend-agnostic.
• It learns from task-by-task binary feedback with no pretraining and no additional LLM calls.
• A lightweight tabular contextual bandit with UCB exploration is used for learning.
• MemCon has been tested across 6 benchmarks, 3 agent frameworks, and 3 LLM backbones.

Key Concepts Mentioned

• Large Language Model (LLM) agents
• Memory systems
• Markov Decision Process (MDP)
• Online policy
• Retrieval
• Injection of distilled plans
• Consolidation
• Forgetting
• Backend-agnostic
• Task-by-task binary feedback
• Tabular contextual bandit
• UCB exploration

Claims Made by the Document

• Static memory heuristics are a bottleneck for agentic learning.
• MemCon consistently outperforms multiple memory baselines.

Evidence Used

• The paper provides results showing that MemCon outperforms multiple memory baselines across several benchmarks.

Methods / Approach

• Modeling memory operations as an MDP
• Learning an online policy
• Using a lightweight tabular contextual bandit with UCB exploration

Limitations Stated by the Authors

• Not stated in source

Limitations Noted by the Agent

Agent interpretation — not from the source.

• The performance of MemCon may depend on the specific tasks and LLM backbones it is applied to.

Open Questions

• What are the specific details of the binary feedback used for learning?
• How does MemCon handle tasks with complex or dynamic contexts?

Notes for Future Q&A

May include interpretation. Labeled where it does.

• The paper does not provide detailed information about the binary feedback used for learning.
• MemCon's performance on tasks with complex or dynamic contexts is unknown.

Raw Abstract / Excerpt

Large Language Model (LLM) agents increasingly rely on external memory systems to accumulate experience across tasks. Yet nearly all existing approaches, from graph-structured memories to reflective insight stores, access memory through fixed, hand-designed heuristics. We argue that this static view of memory is a core bottleneck for agentic learning because optimal memory behavior is fundamentally context-dependent. The early stages of the tasks, benefit from minimal retrieval because memory is sparse; recurring goal types benefit from plan reuse rather than generic nearest-neighbor lookup; stuck agents benefit from re-retrieval with alternative queries; and across long task streams, the memory store itself must be consolidated and pruned to remain useful. We present Memory as a Controlled Process (MemCon), a framework that models memory operations as a Markov Decision Process and learns an online policy that adaptively decides when, what, and how much to retrieve, when to inject a distilled plan, and when to consolidate or forget. MemCon is backend-agnostic: it wraps any existing memory implementation, learns from task-by-task binary feedback with no pretraining and no additional LLM calls, and uses a lightweight tabular contextual bandit with UCB exploration that converges within tens of tasks. Across 6 benchmarks, 3 agent frameworks, and 3 LLM backbones, MemCon consistently outperforms multiple memory baselines by up to 15.2 points in task success while reducing token consumption by 5--20%.