From History to State: Constant-Context Skill Learning for LLM Agents

Researchers propose constant-context skill learning, a framework enabling LLM agents to learn reusable task procedures as lightweight modules rather than storing long prompts in memory. The approach reduces token usage per inference by 2-7x while maintaining or improving performance across multiple benchmark environments, addressing the privacy-capability tradeoff in agent deployment.

This research addresses a fundamental inefficiency in current LLM agent architectures: the need to maintain extensive context windows containing skill descriptions, task histories, and system prompts. By shifting procedural knowledge from prompts into trainable model weights through a context-to-weights framework, the work tackles both computational and privacy concerns that plague production agent deployments.

The technical innovation centers on a deterministic state tracker that compresses task progress into a compact representation, enabling agents to condition solely on current observations and this state block rather than cumulative histories. This architectural shift parallels broader trends in machine learning toward parameter efficiency and inference optimization, where computational burden increasingly determines real-world viability. The use of step-level supervised fine-tuning combined with online reinforcement learning demonstrates practical paths to training such modules effectively.

For the AI industry, this approach directly impacts deployment economics. Cloud-based agents currently expose sensitive intermediate context to external APIs during multi-step workflows, creating privacy vulnerabilities for enterprise and personal assistant applications. Local models gain privacy benefits but sacrifice capability. By reducing token consumption 2-7x without capability loss, constant-context learning makes local deployment more feasible and reduces API costs for cloud systems. The consistent performance across model sizes (4B to 8B parameters) and diverse environments suggests broad applicability.

The trajectory suggests future agent systems will increasingly embed task knowledge as model parameters rather than prompt context. This could reshape how personal assistants balance privacy, cost, and reliability—key factors limiting current adoption. Continued refinement of state representations and scaling to more complex task families will determine whether this approach becomes industry standard.

• →Constant-context skill learning reduces token consumption per inference by 2-7x compared to ReAct baselines while maintaining performance
• →The framework embeds task procedures as trainable model weights rather than storing them in prompts, enabling privacy-preserving local deployment
• →A deterministic state tracker compresses task progress into compact representations, eliminating the need for long context histories
• →Performance matches published agent-training benchmarks across ALFWorld, WebShop, and SciWorld environments with 4B-8B parameter models
• →The approach directly addresses the privacy-cost-capability tension in LLM agent deployment for personal assistants