Memory Papers
Memory Architecture

Mental Model Management: An Operator-Based Framework for LLM Memory

AuthorsOliver Kramer

arXiv 20262026
arXivPDF

TL;DR

Mental Model Management (3M) uses explicit memory and cognitive operators over concept-centered mental models to turn 1,577-word inputs into 1,060-word linked conceptual chunks in a cold-start test.

SharePost on XLinkedIn

Read our summary here, or open the publisher PDF on the next tab.

THE PROBLEM

LLM Memory Lacks Evolving Conceptual Representations

LLM memory systems often accumulate redundant, outdated, or contradictory passages without an explicit mechanism for maintaining compact conceptual representations.

This causes retrieval-augmented agents to reason over scattered text fragments instead of coherent concept-centered models, limiting conflict repair and abstraction.

HOW IT WORKS

Mental Model Management Framework

Mental Model Management (3M) organizes knowledge into Mental Models composed of compact Chunks, transformed by operators like Extract, Add, Update, Merge, and Abstract.

You can think of 3M as a conceptual card catalog: instead of storing whole books, it keeps evolving index cards per concept and rewrites them as understanding improves.

This operator-based design lets 3M maintain compact, coherent conceptual memory that can compress, repair conflicts, and derive new knowledge beyond a static context window.

DIAGRAM

3M Ingestion Control Cycle

This diagram shows how 3M ingests new text using the Extract → Retrieve → operator sequence → Verify control cycle.

DIAGRAM

Behavioral Evaluation Pipeline

This diagram shows how 3M is evaluated via six behavioral test cases with deterministic checks on operator use and redundancy.

PROCESS

How Mental Model Management Handles Text Ingestion

  1. 01

    Extract

    3M uses the Extract operator to identify important concepts and compact chunks from input text T, producing transient conceptual candidates C.

  2. 02

    Retrieve

    3M applies Retrieve to search existing Mental Models in Markdown memory, forming a working subset MC that contextualizes new information.

  3. 03

    Operator Application

    Guided by a controller, 3M invokes operators like Add, Update, Connect, Infer, Conflict Repair, or Abstract to produce a staged state fM.

  4. 04

    Verify

    3M runs Verify to check proposed changes against available knowledge and the source text before committing the updated Mental Models M′.

KEY CONTRIBUTIONS

Key Contributions

  • 01

    Operator Vocabulary for Mental Models

    Mental Model Management (3M) defines operators such as Extract, Retrieve, Add, Update, Merge, Split, Connect, Compress, Prune, and Verify to transform concept-centered Mental Models.

  • 02

    Markdown-Native Reference Manager

    3M is instantiated as a single-process Python manager that stores persistent knowledge exclusively as Markdown files, each capped at 1,500 words per concept.

  • 03

    Executable Behavioral Evaluation Suite

    3M is evaluated on six deterministic behavioral cases, covering eight operator types and tracking redundancy via maximum pairwise token-set Jaccard similarity J with bounds up to 0.478.

RESULTS

By the Numbers

Input length

1,577 words

+517 words over committed memory size

Committed length

1,060 words

67.2% of input length

Relations stored

24 relations

plus 4 knowledge gaps across three models

Max Jaccard overlap

0.478

mean 0.274 across six behavioral cases

The cold-start Evolution Strategies ingestion benchmark tests whether 3M can turn a 1,577-word document into compact, linked Mental Models. The 1,060-word committed memory with 24 relations and four knowledge gaps shows that 3M builds structured conceptual representations instead of storing raw summaries.

BENCHMARK

By the Numbers

The cold-start Evolution Strategies ingestion benchmark tests whether 3M can turn a 1,577-word document into compact, linked Mental Models. The 1,060-word committed memory with 24 relations and four knowledge gaps shows that 3M builds structured conceptual representations instead of storing raw summaries.

BENCHMARK

Cold-Start Evolution Strategies Ingestion

Word count comparison between input text and 3M's committed Mental Models.

KEY INSIGHT

The Counterintuitive Finding

Processing more text with 3M does not necessarily increase memory size; the cold-start run shrank 1,577 words to 1,060 words while adding relations and gaps.

This is surprising because LLM memory systems usually grow monotonically, but 3M uses Update, Merge, Compress, and Prune to reduce redundancy as knowledge accumulates.

WHY IT MATTERS

What this unlocks for the field

Mental Model Management enables LLM agents to maintain evolving, concept-centered Mental Models that can be updated, merged, abstracted, and repaired over time.

Builders can now design agents whose long-term memory improves structurally with use, supporting explicit conflict handling, abstraction hierarchies, and gap-driven knowledge acquisition instead of passive document storage.

~12 min read← Back to papers

Related papers

Memory Architecture

A Control Architecture for Training-Free Memory Use

Yanzhen Lu, Muchen Jiang et al.

· 2026

TAG routes low-confidence steps to uncertainty-based routing, filters them with guarded acceptance with rollback, chooses between bank selection across rule and exemplar memory, and prunes via evidence-based retirement inside a unified control loop. On SVAMP and ASDiv, TAG reaches 81.0% and 85.2% accuracy, improving over the 74.0% and 77.5% no-memory baselines while a compute-matched Retry baseline stays flat.

arXiv:2604.18206Read explainer

Questions about this paper?

Paper: Mental Model Management: An Operator-Based Framework for LLM Memory

Answers use this explainer on Memory Papers.

Checking…