Title: Cognitive Distillation and Schema Formation for Autonomous AI Agents Supervisors: Prof. Okba Tibermacine

This is the short version. The short version is still long. Proceed accordingly.

The problem, filed as a support ticket

Recent LLM-based agents are genuinely impressive until you ask them to learn. They reason, plan, call tools, navigate the web, and interact with their environment. Then, on the next occurrence of the same task, they do all of it again, with the enthusiasm of someone who has never seen the task before.

This is not technically accurate — they have seen it. Every observation, action, and outcome is dutifully appended to an interaction history. The agent does not learn from this history so much as re-read it, because re-reading is simpler than understanding. The approach produces three failure modes, each more expensive than the last:

• Inefficiency. Every decision requires processing an ever-growing backlog. The computational cost of “what should I do next” becomes proportional to “everything that has ever happened.”
• Fragility. Matching is literal, not abstract. Change the color of a button, and the agent no longer recognizes the button, or the task, or possibly the concept of buttons.
• Information Overload. Memory saturates with redundant, noisy, and outdated entries. No mechanism exists to distinguish the valuable from the irrelevant, so the valuable and the irrelevant share a single, very large pile.

Humans solved this problem millennia ago, largely by not filing the whole of yesterday before making breakfast. We intend to copy their homework and cite them.

The proposal

We introduce cognitive distillation: the systematic transformation of raw experience into compact, executable schemas — reusable behavioral structures that function as the agent’s mental playbook.

The project designs, implements, and formally analyzes agents that do this continuously: form schemas, consolidate them, maintain them, and, when necessary, forget them. The end state is a principled framework for autonomous expertise development. The agent will eventually be good at things because it has been doing them, rather than despite it.

Core research question

How can an autonomous AI agent dynamically compress its unstructured experiences into a structured and executable “mental playbook” that maximizes task success while minimizing computational cost?

The answer is not included in this abstract. It is, in a sense, the entire point of the next four years.

Objectives

1. Distillation Trigger Mechanism. Detect when a collection of experiences has become stable and recurrent enough to justify abstraction. The agent should not distill after one incident; neither should it distill after ten thousand.
2. Self-Verifying Behavioral Sandbox. An internal execution environment where candidate behavioral scripts are generated, tested, validated, and refined before being admitted to long-term memory. Effectively, a code review for the agent’s own habits.
3. Adaptive Pruning System. Biologically inspired forgetting: obsolete, redundant, or ineffective schemas decay and are removed. Memory that edits itself is the goal. Memory that only grows is a server with a disk alert.
4. A Theory of Artificial Expertise. A mathematical framework characterizing expertise as a function of task performance, adaptability, knowledge compactness, and computational efficiency. So we can state, in equations, what we suspect in prose.
5. Stability and Convergence Proofs. Formal guarantees that the resulting memory stays bounded, stable, and efficient over time, rather than exhibiting what is technically known as “uncontrolled growth.”

Methodology

Four phases. They are sequential, because phases that are not sequential are called chaos.

Phase 1 — Program Induction. Transform raw interaction histories into compact, executable programs. The agent identifies recurring behavioral patterns and synthesizes reusable procedures from a dynamically evolving library of actions and abstractions. Generated programs are executed and validated in the sandbox before being promoted to memory. Correctness, safety, and effectiveness are checked first. The order of these checks matters.

Phase 2 — Schema Graph Construction. Build a structured Schema Graph: nodes are abstract world states, beliefs, or preconditions; edges are verified executable programs that transform one state into another. This graph is the agent’s long-term semantic memory, and it supports retrieval and reasoning without requiring the agent to re-read its own diary.

Phase 3 — Active Forgetting and Knowledge Consolidation. Inspired by synaptic pruning. Rarely used, overly complex, or poorly performing schemas decay; successful ones are reinforced through use and validation. We are building, in the least comforting phrasing available, an agent that forgets on purpose.

Phase 4 — Real-World Benchmarking. Evaluate in open-ended environments requiring long-term planning and adaptation: WebArena for web navigation, Crafter for survival. Results are compared against retrieval-based memory systems and long-context agents. The baseline is deliberately low, but the comparison is fair.

Expected contributions

• A new memory paradigm for autonomous agents, in which episodic experience becomes structured, executable knowledge.
• Theoretical foundations for artificial expertise, including formal definitions and quantitative measures — groundwork for future lifelong-learning agents.
• Algorithms for schema formation and maintenance, covering induction, validation, consolidation, and forgetting.
• Formal guarantees on memory evolution, including stability, convergence, and resource-performance trade-offs.

References

1. Anokhin, P., et al. (2024). AriGraph: Learning Knowledge Graph World Models with Episodic Memory for LLM Agents. arXiv:2407.04363. https://doi.org/10.48550/arxiv.2407.04363
2. Hu, Y., et al. (2025). Memory in the Age of AI Agents. arXiv:2512.13564. https://doi.org/10.48550/arxiv.2512.13564
3. Hatalis, K., et al. (2024). Memory Matters: The Need to Improve Long-Term Memory in LLM-Agents. AAAI Symposium Series. https://doi.org/10.1609/aaaiss.v2i1.27688
4. Du, P. (2026). Memory for Autonomous LLM Agents: Mechanisms, Evaluation, and Emerging Frontiers.
5. Yang, C., et al. (2026). Graph-based Agent Memory: Taxonomy, Techniques, and Applications. arXiv:2602.05665. https://doi.org/10.48550/arxiv.2602.05665
6. Jia, S., et al. (2025). PISA: A Pragmatic Psych-Inspired Unified Memory System for Enhanced AI Agency. arXiv:2510.15966. https://doi.org/10.48550/arxiv.2510.15966
7. Xu, W., et al. (2025). A-MEM: Agentic Memory for LLM Agents. arXiv:2502.12110. https://doi.org/10.48550/arxiv.2502.12110
8. Abtahi, S. M., et al. (2026). Memanto: Typed Semantic Memory with Information-Theoretic Retrieval for Long-Horizon Agents.
9. Wei, L., et al. (2026). FadeMem: Biologically-Inspired Forgetting for Efficient Agent Memory. arXiv:2601.18642. https://doi.org/10.48550/arxiv.2601.18642
10. Zhang, Z., et al. (2025). A Survey on the Memory Mechanism of Large Language Model based Agents. ACM TOIS. https://doi.org/10.1145/3748302