Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Active Context Compression: Autonomous Memory Management in LLM Agents

Nikhil Verma

TL;DR

This work tackles the context window bottleneck in LLM agents by introducing Focus, an agent-centric mechanism for active intra-trajectory memory compression. Through the Focus Loop (start_focus, explore, complete_focus, withdraw) and a Knowledge block, the agent prunes raw history while preserving essential learnings, creating a sawtooth context pattern. In experiments on five SWE-bench Lite tasks with aggressive prompting, Focus achieves about 22.7% overall token savings without reducing task accuracy, and shows substantial savings on exploration-heavy tasks. The results demonstrate that deliberate prompting and an agent-controlled compression cadence can enable cost-efficient, autonomous memory management for long-horizon software engineering tasks, with opportunities for further refinement via fine-tuning or RL-based approaches.

Abstract

Large Language Model (LLM) agents struggle with long-horizon software engineering tasks due to "Context Bloat." As interaction history grows, computational costs explode, latency increases, and reasoning capabilities degrade due to distraction by irrelevant past errors. Existing solutions often rely on passive, external summarization mechanisms that the agent cannot control. This paper proposes Focus, an agent-centric architecture inspired by the biological exploration strategies of Physarum polycephalum (slime mold). The Focus Agent autonomously decides when to consolidate key learnings into a persistent "Knowledge" block and actively withdraws (prunes) the raw interaction history. Using an optimized scaffold matching industry best practices (persistent bash + string-replacement editor), we evaluated Focus on N=5 context-intensive instances from SWE-bench Lite using Claude Haiku 4.5. With aggressive prompting that encourages frequent compression, Focus achieves 22.7% token reduction (14.9M -> 11.5M tokens) while maintaining identical accuracy (3/5 = 60% for both agents). Focus performed 6.0 autonomous compressions per task on average, with token savings up to 57% on individual instances. We demonstrate that capable models can autonomously self-regulate their context when given appropriate tools and prompting, opening pathways for cost-aware agentic systems without sacrificing task performance.

Active Context Compression: Autonomous Memory Management in LLM Agents

TL;DR

This work tackles the context window bottleneck in LLM agents by introducing Focus, an agent-centric mechanism for active intra-trajectory memory compression. Through the Focus Loop (start_focus, explore, complete_focus, withdraw) and a Knowledge block, the agent prunes raw history while preserving essential learnings, creating a sawtooth context pattern. In experiments on five SWE-bench Lite tasks with aggressive prompting, Focus achieves about 22.7% overall token savings without reducing task accuracy, and shows substantial savings on exploration-heavy tasks. The results demonstrate that deliberate prompting and an agent-controlled compression cadence can enable cost-efficient, autonomous memory management for long-horizon software engineering tasks, with opportunities for further refinement via fine-tuning or RL-based approaches.

Abstract

Large Language Model (LLM) agents struggle with long-horizon software engineering tasks due to "Context Bloat." As interaction history grows, computational costs explode, latency increases, and reasoning capabilities degrade due to distraction by irrelevant past errors. Existing solutions often rely on passive, external summarization mechanisms that the agent cannot control. This paper proposes Focus, an agent-centric architecture inspired by the biological exploration strategies of Physarum polycephalum (slime mold). The Focus Agent autonomously decides when to consolidate key learnings into a persistent "Knowledge" block and actively withdraws (prunes) the raw interaction history. Using an optimized scaffold matching industry best practices (persistent bash + string-replacement editor), we evaluated Focus on N=5 context-intensive instances from SWE-bench Lite using Claude Haiku 4.5. With aggressive prompting that encourages frequent compression, Focus achieves 22.7% token reduction (14.9M -> 11.5M tokens) while maintaining identical accuracy (3/5 = 60% for both agents). Focus performed 6.0 autonomous compressions per task on average, with token savings up to 57% on individual instances. We demonstrate that capable models can autonomously self-regulate their context when given appropriate tools and prompting, opening pathways for cost-aware agentic systems without sacrificing task performance.
Paper Structure (18 sections, 2 figures, 2 tables)

This paper contains 18 sections, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. The Slime Mold Analogy
  5. Architecture: The Focus Loop
  6. Experiments
  7. Optimized Scaffold
  8. Aggressive Compression Prompting
  9. Results
  10. Key Findings
  11. Case Study: Maximum Savings (matplotlib-26020)
  12. Case Study: When Compression Adds Overhead (pylint-7080)
  13. Analysis
  14. Discussion
  15. The Cognitive Tax of Compression
  16. ...and 3 more sections

Figures (2)

  • Figure 1: Total token consumption across 5 hard instances. Focus reduces usage by 22.7% through aggressive model-controlled compression while maintaining identical accuracy.
  • Figure 2: Conceptual sawtooth pattern of context growth. Focus (blue) exhibits periodic compressions (drops) while Baseline (red) grows monotonically. With aggressive prompting, Focus compresses every 10-15 tool calls, preventing context bloat while preserving learnings in a persistent Knowledge block.