Criticity-Weighted Memory Regulation for Long-Running LLM Agents

Conceptual research proposal on memory regulation for long-running LLM agents.

Architecture overview

LLM
↓
Latent Memory Index
↓
External Episodic Storage
↓
Criticity-Weighted Memory Regulation

Live version of the proposal:

Live version of the proposal:
https://frthierry09.github.io/Criticity-Weighted-Memory-Regulation/

Overview

Long-running LLM agents tend to accumulate noise, errors, and irrelevant context in memory.
Over time this degrades reasoning stability and coherence.

This repository presents a conceptual framework called Criticity-Weighted Memory Regulation, where stored memories are dynamically weighted according to their importance (“criticality”).

High-value information persists, while low-signal or erroneous context gradually fades.

Core Idea

Instead of treating all memories equally, the system assigns a criticality score based on factors such as:

• relevance to the task
• reliability of the information
• impact on future reasoning

Memories with higher criticality remain stable in the agent’s working context, while low-criticality memories decay or are filtered.

Related Research Areas

• Continual learning for LLM agents
• Agent memory architectures
• Vector database regulation
• Long-running autonomous agents
• Cognitive architectures for AI

Potential Benefits

• reduced accumulation of noise in long conversations
• improved reasoning stability
• automatic filtering of low-value context
• better long-term coherence in agent behavior

Paper

The full proposal is available here:

[Read the full proposal] (index.html)

Status

Conceptual research proposal open for discussion.

Author

François Thierry

Keywords: LLM agents, agent memory, continual learning, memory regulation, vector database, AI architecture

#AI #LLM #AgentMemory #ContinualLearning #MachineLearning #AIResearch