mlsanityData to 7 Oct 2026

ACON: Optimizing Context Compression for Long-horizon LLM Agents

arXiv 2510.00615 ↗First posted 1 Oct 2025cs.AI cs.CLcited by 126

On the map: LLM agents and tool use → Agent memory systems

also in 1 more
Abstract (arXiv)

Large language models (LLMs) are increasingly deployed as agents in dynamic real-world environments, where success depends on maintaining precise records of actions and observations. However, the resulting unbounded context growth in long-horizon agentic tasks makes two critical bottlenecks: prohibitive inference memory costs and reasoning degradation due to irrelevant information. Existing compression methods fail to fully address this, often relying on brittle heuristics or requiring parameter updates impractical for proprietary or large-scale LLMs. We introduce Agent Context Optimization (ACON), a unified framework that optimally compresses both observations and history into concise, informative representations. Distinct from prior works, ACON employs an optimization in natural language space: it iteratively refines compression guidelines based on failure analysis of the agent, ensuring critical state information is preserved without model fine-tuning. To further minimize computational overhead, we distill the optimized compressor into smaller models. Experiments on AppWorld, OfficeBench, and Multi-objective QA demonstrate that ACON reduces peak token usage by 26-54% while improving task success over existing compression baselines. Notably, it enables smaller LMs to function effectively as long-horizon agents, achieving up to 46% performance improvement by mitigating context distraction. Our code is available at https://github.com/microsoft/acon.

Key facts. Cited by 126 papers in the map (data through 7 Oct 2026). Placed in Agent memory systems, one of 16 directions in LLM agents and tool use. Builds on 20 of its 61 references that are in the map. First posted on arXiv 1 Oct 2025.

Builds on

20 of 61 references.

Closest ancestors · 4

Same direction, newest first

  1. 2025ReSum: Unlocking Long-Horizon Search Intelligence via Context Summarization16 Sep
  2. 2025MEM1: Learning to Synergize Memory and Reasoning for Efficient Long-Horizon Agents18 Jun
  3. 2025A-MEM: Agentic Memory for LLM Agents17 Feb
  4. 2024Evaluating Very Long-Term Conversational Memory of LLM Agents27 Feb

Other references · 16

  1. 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Modelscited by 8,851
  2. 2025Qwen3 Technical Reportcited by 8,195
  3. 2025DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learningcited by 6,146
  4. 2025Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilitiescited by 3,655
  5. 2025Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learningcited by 1,371
  6. 2024Phi-4 Technical Reportcited by 675
  7. 2025BrowseComp: A Simple Yet Challenging Benchmark for Browsing Agentscited by 616
  8. 2025Kimi K2: Open Agentic Intelligencecited by 523
  9. 2025MemAgent: Reshaping Long-Context LLM with Multi-Conv RL-based Memory Agentcited by 235
  10. 2024Windows Agent Arena: Evaluating Multi-Modal OS Agents at Scalecited by 212
  11. 2025Measuring AI Ability to Complete Long Software Taskscited by 142
  12. 2025LightThinker: Thinking Step-by-Step Compressioncited by 104
4 more
  1. 2024OfficeBench: Benchmarking Language Agents across Multiple Applications for Office Automationcited by 66
  2. 2025OdysseyBench: Evaluating LLM Agents on Long-Horizon Complex Office Application Workflowscited by 54
  3. 2025EpiCache: Episodic KV Cache Management for Long-Term Conversation on Resource-Constrained Environmentscited by 7
  4. 2025Enhancing Reasoning Capabilities of Small Language Models with Blueprints and Prompt Template Searchcited by 3

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2025: 6; 2026: 120.

126 papers.

Sort:

By area:

  1. 2025DeepAgent: A General Reasoning Agent with Scalable Toolsetscited by 69
  2. 2026Meta Context Engineering via Agentic Skill Evolutioncited by 42
  3. 2026SWE-Pruner: Self-Adaptive Context Pruning for Coding Agentscited by 40
  4. 2026A Survey of Agent Memory in the Second Half: Towards Self-Evolving and Long-Horizon Agentscited by 40
  5. 2025Context as a Tool: Context Management for Long-Horizon SWE-Agentscited by 38
  6. 2026The Long-Horizon Task Mirage? Diagnosing Where and Why Agentic Systems Breakcited by 25
  7. 2026AgentOCR: Reimagining Agent History via Optical Self-Compressioncited by 22
  8. 2026Anatomy of Agentic Memory: Taxonomy and Empirical Analysis of Evaluation and System Limitationscited by 20
  9. 2026Building Effective AI Coding Agents for the Terminal: Scaffolding, Harness, Context Engineering, and Lessons Learnedcited by 18
  10. 2026A Self-Evolving Framework for Efficient Terminal Agents via Observational Context Compressioncited by 12
  11. 2026The Evolution of Tool Use in LLM Agents: From Single-Tool Call to Multi-Tool Orchestrationcited by 11
  12. 2026VitaBench 2.0: Evaluating Personalized and Proactive Agents in Long-Term User Interactionscited by 11
38 more
  1. 2026SAM: State-Adaptive Memory for Long-Horizon Reasoning Agentcited by 10
  2. 2026Do Not Waste Your Rollouts: Recycling Search Experience for Efficient Test-Time Scalingcited by 9
  3. 2026Group of Skills: Group-Structured Skill Retrieval for Agent Skill Librariescited by 9
  4. 2026From Question Answering to Task Completion: A Survey on Agent System and Harness Designcited by 9
  5. 2025ContextPilot: Fast Long-Context Inference via Context Reusecited by 8
  6. 2026The Missing Memory Hierarchy: Demand Paging for LLM Context Windowscited by 8
  7. 2026Agent-BRACE: Decoupling Beliefs from Actions in Long-Horizon Tasks via Verbalized State Uncertaintycited by 8
  8. 2026Self-GC: Self-Governing Context for Long-Horizon LLM Agentscited by 8
  9. 2026ACM: Agentic Context Management for Long Horizon Taskscited by 8
  10. 2026OCR-Memory: Optical Context Retrieval for Long-Horizon Agent Memorycited by 7
  11. 2026Better Harnesses, Smaller Models: Building 90% Cheaper Agents via Automated Harness Adaptationcited by 7
  12. 2025FoldAct: Efficient and Stable Context Folding for Long-Horizon Search Agentscited by 6
  13. 2026MemCollab: Cross-Model Memory Collaboration via Contrastive Trajectory Distillationcited by 6
  14. 2025PRInTS: Reward Modeling for Long-Horizon Information Seekingcited by 5
  15. 2026AgentSPEX: An Agent SPecification and EXecution Languagecited by 5
  16. 2026ContextEcho: A Benchmark for Persona Drift in Long Agentic-Coding Sessionscited by 5
  17. 2026CoMem: Context Management with A Decoupled Long-Context Modelcited by 5
  18. 2026Governance Decay: How Context Compaction Silently Erases Safety Constraints in Long-Horizon LLM Agentscited by 5
  19. 2026Toward Reliable Context Compression for Long-Horizon Agents: An Empirical Study of Execution Instabilitycited by 5
  20. 2025A Benchmark for Procedural Memory Retrieval in Language Agentscited by 4
  21. 2026Optimizing Agentic Workflows using Meta-toolscited by 4
  22. 2026ContextWeaver: Selective and Dependency-Structured Memory Construction for LLM Agentscited by 4
  23. 2026PEEK: Context Map as an Orientation Cache for Long-Context LLM Agentscited by 4
  24. 2026Parallel Context Compaction for Long-Horizon LLM Agent Servingcited by 4
  25. 2026Learning What to Remember: Observability-Safe Memory Retention via Constrained Optimization for Long-Horizon Language Agentscited by 4
  26. 2026Learning What Not to Forget: Long-Horizon Agent Memory from a Few Kilobytes of Learningcited by 4
  27. 2026CoACT: Action-Preserving Observation Compression for Coding Agentscited by 4
  28. 2026What to Keep, What to Forget: A Rate--Distortion View of Memory Compaction in LLMs and Agentscited by 4
  29. 2026MemRouter: Memory-as-Embedding Routing for Long-Term Conversational Agentscited by 3
  30. 2026Organize then Retrieve: Hierarchical Memory Navigation for Efficient Agentscited by 3
  31. 2026LLM Agents Are Latent Context Managers: Eliciting Self-Managed Context via State Proprioceptioncited by 3
  32. 2026ECHO: Prune To Act, Trace To Learn With Selective Turn Memory In Agentic RLcited by 3
  33. 2026Context Pruning for Coding Agents via Multi-Rubric Latent Reasoningcited by 2
  34. 2026EvoDS: Self-Evolving Autonomous Data Science Agent with Skill Learning and Context Managementcited by 2
  35. 2026Less Context, Better Agents: Efficient Context Engineering for Long-Horizon Tool-Using LLM Agentscited by 2
  36. 2026When Summaries Distort Decisions: Information Fidelity in LLM-Compressed Financial Analysiscited by 2
  37. 2026ACE: Pluggable Adaptive Context Elasticizer across Agentscited by 2
  38. 2026SWE-Pruner Pro: The Coder LLM Already Knows What to Prunecited by 2

50 most cited of 126 All 126 (JSON)

Similar, not cited

Close in wording, no citation either way

  1. 2025Can Compressed LLMs Truly Act? An Empirical Evaluation of Agentic Capabilities in LLM Compression
  2. 2026AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications
  3. 2026PRO-LONG: Programmatic Memory Enables Long-Horizon Reasoning
  4. 2026MemRefine: LLM-Guided Compression for Long-Term Agent Memory
  5. 2026Memory for Autonomous LLM Agents:Mechanisms, Evaluation, and Emerging Frontiers
  6. 2026Escaping the Context Bottleneck: Active Context Curation for LLM Agents via Reinforcement Learning
  7. 2026AgentCompress: Task-Aware Compression for Affordable Large Language Model Agents
  8. 2025Scaling Long-Horizon LLM Agent via Context-Folding