Plans Don't Persist: Why Context Management Is Load Bearing for LLM Agents
Summary
This paper investigates how LLM agents lose plan information as it gets evicted from context during long interactions. Using replay pairing and compression stress tests, the authors show that standard agents do not carry plans as persistent state, and propose diagnostics to measure plan signal decay.
View Cached Full Text
Cached at: 06/25/26, 05:13 PM
Paper page - Plans Don’t Persist: Why Context Management Is Load Bearing for LLM Agents
Source: https://huggingface.co/papers/2606.22953
Abstract
Standard LLM agents rely on plan content remaining in context rather than maintaining it as persistent state, with evidence shown through replay pairing diagnostics and compression stress tests.
Long-horizon agents depend oncontext management: systems compress, summarize, and evict old tokens so tasks can continue beyond finite windows. That is safe only when dropped information is no longer needed or has been internalized. Plans are the stress case: they are written early, used for many steps, and first to be evicted. We introduce replay pairing, a diagnostic that runs the same trajectory with and without the plan in history and measureshidden-state cosine distance. On Llama-3.1-70B,plan signalspikes to 0.453 one step after the plan, then falls 4.1x in a single action-observation step; HotpotQA falls 12.4x. This is evidence that standard LLM agents do not carry plans forward as persistent state, and instead depend on the plan remaining in context. Alayer-L32 probedetects this decay as a diagnostic, not as proof that it reads plan content itself. Reasoning models add a measurement confound: their `<think>` traces re-derive plan content, so standard stripping leaves plan evidence in the stripped condition. We name this thereasoning-trace confoundand fix it withstrict stripping, which removes prior `<think>` blocks from the stripped run only. It recovers +163% of the step+1 signal in-sample and +153% held out, while not meaningfully changing non-reasoning Llama (+4.8%). On DeepSeek-R1-Distill-Llama-70B, a Llama-trained probe transfers atAUROC0.748 (p=6e-4), while R1-specific probes reach 1.000, suggesting R1 encodesplan signalin a different hidden-state direction. Finally, a compression stress test shows the practical cost: naive plan eviction cutsALFWorldsuccess by 34.7pp, while probe-gated re-surfacing does not recover it. The contribution is a measurement and stress-test framework showing that agent-critical information can be context-resident rather than persistent.Context managementis load bearing, but plan protection alone is not enough.
View arXiv pageView PDFAdd to collection
Get this paper in your agent:
hf papers read 2606\.22953
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2606.22953 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2606.22953 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2606.22953 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
Less Context, Better Agents: Efficient Context Engineering for Long-Horizon Tool-Using LLM Agents
This paper evaluates context engineering configurations for LLM agents in enterprise tool-use workflows, showing that summarization with selective pruning achieves 91.6% accuracy while reducing token usage by over 60% compared to full-context baselines.
@omarsar0: // The Memory Curse in LLM Agents // (bookmark it) Long histories apparently degrades agents as they become increasingl…
This research paper identifies the 'memory curse' in LLM agents, demonstrating that expanded context windows systematically degrade cooperative behavior in multi-agent social dilemmas by eroding forward-looking intent. The authors show that targeted fine-tuning, synthetic memory sanitization, and reducing explicit Chain-of-Thought reasoning can effectively mitigate this behavioral decay.
Toward Reliable Context Compression for Long-Horizon Agents: An Empirical Study of Execution Instability
This empirical study investigates how recurrent context compression affects long-horizon agent behavior, showing that compression can weaken recent interaction influence and cause instability. The authors introduce TRACE, a verifier-guided framework that improves compression reliability and performance on AppWorld.
When Tools Fail: Benchmarking Dynamic Replanning and Anomaly Recovery in LLM Agents
The ToolMaze benchmark evaluates LLM agents' ability to handle real-world tool failures, revealing that implicit semantic failures cause the largest performance drops and that dynamic replanning remains a critical bottleneck not addressed by scaling or prompting.
Why Retrying Fails: Context Contamination in LLM Agent Pipelines
This paper introduces the Context-Contaminated Restart Model (CCRM) to formally analyze how failed attempts in LLM agent pipelines contaminate context and increase error rates during retries. It provides theoretical proofs and validates the model against SWE-bench data, showing significant discrepancies with standard independent models.