Plans Don't Persist: Why Context Management Is Load Bearing for LLM Agents

Abstract: Long-horizon agents depend on context 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 measures hidden-state cosine distance. On Llama-3.1-70B, plan signal spikes 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. A layer-L32 probe detects this decay as a diagnostic, not as proof that it reads plan content itself. Reasoning models add a measurement confound: their > traces re-derive plan content, so standard stripping leaves plan evidence in the stripped condition. We name this the reasoning-trace confound and fix it with strict 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 at AUROC 0.748 (p=6e-4), while R1-specific probes reach 1.000, suggesting R1 encodes plan signal in a different hidden-state direction. Finally, a compression stress test shows the practical cost: naive plan eviction cuts ALFWorld success 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 management is load bearing, but plan protection alone is not enough.

• The paper finds that LLM agent plan signals decay sharply after insertion, indicating that plans exist only as ephemeral context artifacts.
• Methodology leverages replay pairing and strict stripping protocols to isolate and measure plan signals across varied LLM architectures.
• Empirical results reveal that naive plan eviction reduces task success substantially, emphasizing context management as a critical bottleneck.

Plans Do Not Persist: Context Management as a Critical Bottleneck for LLM Agents

Introduction

This paper presents a rigorous representational analysis of plan persistence in LLM agents and investigates the implications of context compression and management. Through empirical studies on Llama-3.1-70B and related variants, the authors demonstrate that the plan-related signal in LLM hidden states decays precipitously within a single action-observation cycle, indicating that such plans exist as context-time artifacts rather than persistent internal states. The methodological innovations include the replay pairing diagnostic, strict stripping protocols to address reasoning-trace contamination, and cross-domain/model probe-transfer validation. The results underscore that context management, specifically the handling of agent-critical tokens like plans, is a load-bearing component underlying reliable agent behavior and that naive plan protection, even when aided by representational probes, does not rescue performance under aggressive compression.

Method: Replay Pairing and Plan Signal Measurement

The centerpiece of the empirical analysis is the replay pairing protocol Figure 1. Two matched agent trajectories are executed: one with the plan present in history (condition A), and one with the plan removed (condition B). Hidden states $\mathbf{h}_A$ and $\mathbf{h}_B$ are collected at every step and layer, and the plan signal is defined as the per-step cosine distance between these hidden states. This diagnostic isolates the representational footprint of the plan.

Figure 1: Replay pairing schematic illustrating the plan signal as the cosine distance between two matched trajectories, showing plan re-read from context rather than persistent internal state.

Strict stripping is introduced for reasoning models that emit > traces. Standard replay pairing is confounded by reasoning traces that re-derive plan content, contaminating the supposedly plan-free condition. Strict stripping removes these traces only from B, recovering the masked plan signal and validating the protocol on non-reasoning models as a verified no-op.

Empirical Findings: Plan Signal Decay

The core finding is that in Llama-3.1-70B ReAct agents, plan signal spikes immediately after plan insertion ($0.453$ at step +1) but decays $4.1\times$ within a single cycle, stabilizing near zero ($\sim 0.027$ at step +5). HotpotQA trajectories exhibit even more rapid decay ($12.4\times$). Entirety of the plan's influence is context-resident; once the plan exits the window, behaviorally aligned actions cease even though hidden-state signals fade quickly.

Figure 2: Plan signal decay curve across steps, with rapid drop-off one step after plan injection.

Layer-wise profiles indicate that plan signal is sharply localized, peaking at $L_{32}$ in Llama-70B, with the decay dynamics replicated across all six ALFWorld task types. Cross-domain and cross-scale validation (Llama-8B) confirms the generality of the findings: the plan signal shape and decay are invariant to architecture and task complexity.

Figure 3: Layer profile at step+1 showing plan signal localization at mid-network depth ($L_{32}$).

Figure 4: Multi-model comparison confirming similar decay behavior across model scales.

Diagnostic Probe and Detection Performance

A Ridge regression probe trained on peak layer hidden states ($L_{32}$) achieves $R^2=0.875$ and binary AUROC $\mathbf{h}_B$0 (ALFWorld), with zero-shot transfer to HotpotQA (AUROC $\mathbf{h}_B$1). Although step-index leakage inflates AUROC, mixed-effects analysis and graded probe tasks suggest non-trivial content-specific signal. Importantly, the probe's early warning capability allows detection of plan-signal decay up to $\mathbf{h}_B$2 steps before behavioral deviation.

Figure 5: Step-index leakage analysis demonstrates necessity of content-specific controls for probe validity.

Figure 6: Reliability diagram shows near-perfect calibration of the $\mathbf{h}_B$3 plan presence probe.

Reasoning Models: Trace Contamination and Directional Encoding

Reasoning models such as DeepSeek-R1-Distill-Llama-70B emit <think> traces, leading to measurement confounds. Standard replay pairing undercounts plan signal, but strict stripping recovers $\mathbf{h}_B$4 of the lost signal. Probe-transfer studies show that reasoning models encode the plan-related signal in a direction nearly orthogonal (angle $\mathbf{h}_B$5) to the standard Llama direction. Qwen3-native (enable_thinking=True) displays persistent plan-conditional drift, evidencing model-family-dependent encoding.

Compression Stress Test: Operational Failure Under Plan Eviction

A context-compression stress test on ALFWorld Figure 7 reveals catastrophic loss of task success: naive eviction of plan tokens reduces success by $\mathbf{h}_B$6 percentage points. Neither plan-protection nor probe-gated plan resurfacing recovers performance, indicating that plan persistence alone is insufficient for agent reliability under aggressive context evictions. The plan's existential safety hinges on its presence in context rather than internalization in hidden state.

Figure 7: Compression stress test confirming unsafeness of naive plan eviction and futility of plan-aware policies at fixed context budget.

Discussion and Implications

The results redefine the operational boundaries of context management for agentic LLMs. Critical information such as plans, constraints, and tool schemas may be irretrievably context-resident, making compression, summarization, and eviction protocols fundamentally unsafe without granular representational diagnostics. Reasoning models add subtlety: their self-refreshing traces shift encoding, but this does not equate to state persistence. Probe-triggered interventions cannot replace context memory.

Theoretically, the findings challenge assumptions in agent frameworks like ReAct, Chain-of-Thought, and Reflexion, which presume explicit plans can be evicted once “internalized.” Absent durable state encoding, plans are not absorbed, but continually re-read. This suggests open research directions in designing architectures or training paradigms that foster genuinely persistent internal memory for agent-critical instructions.

Conclusion

The paper establishes that in current LLM architectures, plans remain context-time entities with representational imprints that decay rapidly and are not stored persistently in hidden state. Context management is thus a critical bottleneck: evicting plans (or other agent-critical tokens) without an accompanying durable encoding sharply degrades agent performance. Probe-based diagnostics illuminate context safety but cannot themselves rescue agent behavior under extreme compression. Practical deployment of LLM agents must treat context management as a safety-critical axis, and future research should focus on developing memory mechanisms that enable true persistence of plans and instructions in agent internal states.