mlsanityData to 14 Oct 2026

When Can Agents Forget Their Reasoning? ICLR for Long-Horizon Agent Context Compression

arXiv 2609.29875 ↗First posted 24 Sep 2026cs.AI cs.CVcited by 0

On the map: LLM reasoning → Reasoning for agents, planning and code also fits: Test-time compute and efficient thinking

also in 2 more
Abstract (arXiv)

Long horizon language model agents continually accumulate reasoning history, increasing context length and inference cost even after earlier decisions have been executed and observed. Unlike static Chain of Thought compression, removing historical reasoning can change future actions and the resulting interaction trajectory. We study when such reasoning can be safely forgotten. We propose Interaction Aware Compression for Long Horizon Reasoning (ICLR), a training free online method that ranks reasoning blocks using frozen proxy entropy while preserving actions, tool calls, and observations. On 260 WorkBuddyBench tasks, ICLR improves average reward from 0.699 to 0.718, while reducing input, output, and cache read tokens by 25.5%, 14.4%, and 33.3%, respectively. Ablations reveal trajectory amplification, where local reasoning deletion produces nonlinear changes in total computation by altering subsequent interaction. Representation probing, activation patching, and controlled trajectory analyses further suggest that historical reasoning becomes more replaceable once task relevant derived state has been reliably externalized into code, files, tool outputs, or environmental feedback. These results characterize agent reasoning as dynamic working state rather than permanent interaction history.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Reasoning for agents, planning and code, one of 17 directions in LLM reasoning. Builds on 24 of its 50 references that are in the map. First posted on arXiv 24 Sep 2026.

Builds on

24 of 50 references.

Closest ancestors · 2

Same direction, newest first

  1. 2026SAM: State-Adaptive Memory for Long-Horizon Reasoning Agent23 May
  2. 2026Active Context Compression: Autonomous Memory Management in LLM Agents12 Jan

Other references · 22

  1. 2022ReAct: Synergizing Reasoning and Acting in Language Modelscited by 8,695
  2. 2022Self-Consistency Improves Chain of Thought Reasoning in Language Modelscited by 5,708
  3. 2023MemGPT: Towards LLMs as Operating Systemscited by 1,389
  4. 2016Understanding intermediate layers using linear classifier probescited by 1,353
  5. 2024Training Large Language Models to Reason in a Continuous Latent Spacecited by 797
  6. 2022Discovering Latent Knowledge in Language Models Without Supervisioncited by 767
  7. 2024Quiet-STaR: Language Models Can Teach Themselves to Think Before Speakingcited by 301
  8. 2024Leave No Context Behind: Efficient Infinite Context Transformers with Infini-attentioncited by 200
  9. 2025Scaling Long-Horizon LLM Agent via Context-Foldingcited by 148
  10. 2025ReSum: Unlocking Long-Horizon Search Intelligence via Context Summarizationcited by 141
  11. 2025ACON: Optimizing Context Compression for Long-horizon LLM Agentscited by 136
  12. 2025AgentFold: Long-Horizon Web Agents with Proactive Context Managementcited by 103
10 more
  1. 2026SWE-Pruner: Self-Adaptive Context Pruning for Coding Agentscited by 44
  2. 2026Beyond Static Summarization: Proactive Memory Extraction for LLM Agentscited by 21
  3. 2026ContextBudget: Budget-Aware Context Management for Long-Horizon Search Agentscited by 14
  4. 2026Self-Compacting Language Model Agentscited by 11
  5. 2026Self-GC: Self-Governing Context for Long-Horizon LLM Agentscited by 11
  6. 2026ACM: Agentic Context Management for Long Horizon Taskscited by 10
  7. 2026Beyond Compaction: Structured Context Eviction for Long-Horizon Agentscited by 8
  8. 2026Chain Of Thought Compression: A Theoretical Analysiscited by 5
  9. 2026CoMem: Context Management with A Decoupled Long-Context Modelcited by 5
  10. 2026Context as an Environment: Programmatic Context Management for Long-Horizon Agentscited by 4

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026StateComp: Learning When to Compress History in Long Horizon Agents
  2. 2026PRO-LONG: Programmatic Memory Enables Long-Horizon Reasoning
  3. 2026Memory Control Signals Emerge Before Action in Long Horizon Agents
  4. 2026FOCUS: Training-Free Decision-Preserving Context Compression for LLM Agents
  5. 2026Beyond Memory: Harnessing Long-Horizon Agents with Explicit Belief States
  6. 2026AMBER: Training Long-Horizon Web Agents through Append-Only Memory
  7. 2026MemWeaver: Weaving Hybrid Memories for Traceable Long-Horizon Agentic Reasoning
  8. 2026AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications