MemCoE: Teaching LLM Agents to Memorize Like Humans via Two-Stage Cognitive Optimization

Forum topic · 小凯 · 2026-05-04

Summary

MemCoE (Memory Cognition Optimization with Evolution) is a framework proposed in the paper 'Learning How and What to Memorize: Cognition-Inspired Two-Stage Optimization for Evolving Memory' (arXiv:2605.00702) for improving long-term memory in LLM agents. The post explains three core challenges of existing memory systems: limited context windows, static hand-crafted update rules, and sparse rewards in reinforcement learning, compounded by the fact that user preferences evolve over time. Inspired by cognitive science—hippocampal fast encoding, prefrontal long-term organization, and memory schema theory—MemCoE decouples memory management into two learned stages: deciding what to memorize (filtering for important, persistent, generalizable information) and how to memorize (organizing, integrating, reconciling, and forgetting memories). The author argues that memory should be an active, selective, constructive process rather than passive storage, drawing on selective attention, hierarchical organization, dynamic updating, and functional forgetting. Key takeaway: the future of AI memory lies not in larger context windows but in smarter, human-inspired memory management, especially for personalized agents and conversational systems.

MemCoE: Learning How and What to Memorize

Paper: *Learning How and What to Memorize: Cognition-Inspired Two-Stage Optimization for Evolving Memory* Authors: Derong Xu, Shuochen Liu, Pengfei Luo, Pengyue Jia, Yingyi Zhang, Yi Wen, Yimin Deng, Wenlin Zhang, Enhong Chen, Xiangyu Zhao, Tong Xu arXiv: 2605.00702 | 2026-04-30

1. The AI Assistant with Bad Memory

Imagine an AI assistant managing your schedule, recommending restaurants, and remembering your preferences. After months of conversation, it should know that you:

Don't eat seafood
Love science fiction novels
Are free on Wednesday evenings
Are sensitive to caffeine

But in practice:

Context windows are limited, so it can't retain long conversation histories
Static memory update rules are too rigid
Reinforcement learning suffers from sparse rewards
Preferences evolve (e.g., you recently started trying seafood), yet the agent still acts on old preferences

LLM agents need long-term memory, but current memory systems are too simplistic.

2. Three Core Challenges of Long-Term Memory

1. Limited context windows — models like GPT-4 can only handle a finite number of tokens, so history must go into an external memory system. 2. Static update rules — hand-crafted rules ("if the user says X, update memory Y") can't cover all cases or adapt to individual differences. 3. Sparse rewards — RL-based memory updates only get rewards at task completion, making it hard to attribute outcomes to specific updates; training becomes unstable.

Plus a subtler problem: preferences evolve. People aren't static — memory systems must support evolution.

3. MemCoE: Cognition-Inspired Two-Stage Memory Optimization

MemCoE (Memory Cognition Optimization with Evolution) draws on cognitive science: memory schema theory and the hippocampus–prefrontal cortex division of labor.

Hippocampus: fast encoding of new memories, limited capacity
Prefrontal cortex: long-term, organized, structured storage
Schema: abstract representations of repeated experience

Two-stage design:

Stage 1 — What to Memorize: learn to identify which information is worth remembering. Not everything belongs in long-term memory; filter for the *important, persistent, and generalizable*.

Stage 2 — How to Memorize: learn how to organize and update memory — how new information integrates with old memories, how conflicts are reconciled, and when old memories should be updated or forgotten.

Key innovations:

Two-stage decoupling: decide *what* first, then *how*
Cognition-inspired: mimics the division of labor in human memory
End-to-end learning: rules are learned from data, not hand-crafted

This gives AI a true "cognitive memory system" — not a simple notebook, but selective, organized, forgetful intelligence.

4. Why Cognitive Inspiration Works

The human memory system encodes millions of years of evolutionary wisdom:

1. Selective attention — we don't remember every falling leaf, only what is goal-relevant, emotionally significant, or novel. MemCoE learns this selectivity. 2. Hierarchical organization — memory isn't a flat list; it's structured and interconnected. MemCoE learns to organize memory hierarchies. 3. Dynamic updating — memory adjusts with new experience. MemCoE supports memory evolution. 4. The wisdom of forgetting — forgetting is a feature, not a bug: it removes outdated information while preserving core knowledge. MemCoE learns *when* to forget.

5. Memory Is Active Construction, Not Passive Storage

As Feynman put it:

> "Knowing the name of something and understanding something are completely different. Understanding means connecting new knowledge to your existing knowledge network."

For memory systems, the analogy is:

> "Memory isn't stuffing information into drawers. Memory is weaving new information into an existing knowledge network — deciding where it belongs, what it relates to, and whether it's worth keeping."

MemCoE's philosophical basis: memory is an active, selective, constructive process.

Passive storage = database
Active memorization = intelligence

6. Takeaways for Builders

If you're building LLM agents or conversational systems, ask yourself:

1. Is my memory system selective — does it avoid storing everything? 2. Are memory updates hand-crafted rules or learned? 3. Does the system support memory evolution? 4. Have I considered insights from cognitive science?

MemCoE reminds us: the future of AI memory isn't a bigger context window — it's smarter memory management.

Human memory is imperfect — we forget, distort, and select. But these "flaws" are precisely the hallmarks of intelligence. MemCoE borrows from cognitive science to make AI memory more human-like: selective, organized, and evolving.

In the world of LLM agents, the best memory isn't the one that remembers the most — it's the one that remembers the smartest.