Can Agents Keep Up with Time? An Empirical Analysis of Temporal Adaptation After a Conversational Hiatus

Hongguan Su, Ming Wang, Shun Zhang, Yifan Wang, Rong Yin, Zhiming Ma, Mengzheng Yang, Peng ChenarXiv 2026 · arXiv

TL;DR

Even when LLM agents are smart enough to check the clock, they often fail to realize that time has passed for the tasks they are managing. This paper identifies a "knowing-doing dissociation" where agents correctly reason that a deadline has passed in their internal chain-of-thought, yet still provide obsolete advice to the user as if the event were in the future.

What Problem Does This Paper Address?

As we move toward persistent AI assistants that manage our lives over weeks or months, a new challenge emerges: the conversational hiatus. A user might book a medical appointment in March for May, stop talking to the agent, and then return in June.

While the agent’s memory of the "May 6th appointment" remains factually correct, the status of that memory has shifted from a future plan to a past event. Existing benchmarks focus on whether agents can remember facts across long horizons, but they don't test temporal state perception and use—the ability to perceive the current time and dynamically re-interpret static records based on that time.

Figure 1: A session-rekindling scenario where the scheduled May 6 health check has passed by the June 5 follow-up, con- trasting stale preparation advice with an appropriate clarify- ing response.

Without this, an agent becomes a "stale" assistant, telling you to fast for a medical checkup that happened a month ago.

Methods

The authors formalize this as a session-rekindling task. They designed controlled service episodes where a historical dialogue and preparation notes are fixed, but the "virtual clock" is advanced to different offsets (e.g., 1 day before the event vs. 30 days after).

To pinpoint exactly where agents fail, they tested five conditions of temporal information:

  1. Natural: The agent must proactively call a Clock tool to find the date.
  2. Delivered: The current date is provided directly in the user prompt.
  3. Forced: A completed clock call is injected into the model's context.
  4. Relation: The agent is explicitly told the relation (e.g., "The event date has passed").
  5. User-reference: The user asks the agent to answer "as of" a specific date.

The evaluation uses a "dual-judge" protocol (automated judges with human-validated rubrics) to measure several observables: tool acquisition ($F$), past-date recognition in reasoning ($R$), and "stale framing" in the final response ($S$).

Figure 3: Evaluation framework showing the fixed episode pipeline, the five temporal-information conditions around the base probe, and the response-level observables.

Datasets

The researchers introduced the Session-Rekindling Benchmark, which includes:

  • 940 main responses collected across five state-of-the-art model routes (including DeepSeek and Qwen variants).
  • 14–19 core service event types (e.g., visa interviews, thesis defenses, medical exams).
  • 350 standalone baseline items to verify basic calendar arithmetic.

The dataset is synthetic, written in Chinese, and designed to isolate the behavioral impact of elapsed time by keeping all other textual inputs byte-identical across evaluation dates.

Key Insights and Ideas

  • The Knowing-Doing Dissociation: This is the most striking finding. In 116 out of 120 "stale" responses from certain models, the agent's own reasoning trace explicitly stated that the date had passed. Yet, the final output still told the user to "prepare for tomorrow's appointment."
  • Deictic Re-anchoring Failure: Agents struggle to shift the "center" of relative time. In their memory, "tonight" and "tomorrow" are bound to the original booking date. When they "rekindle" the session, they fail to re-bind these terms to the actual current night.
  • Helpfulness vs. Accuracy: The authors suggest that "instruction-tuned inertia" might be at play. Models are so conditioned to be "helpful" by following stored checklists that they prioritize providing the requested plan over the "unhelpful" (but correct) realization that the plan is now moot.
  • Information Prominence Matters: Simply knowing the date (Delivered/Forced) isn't enough to fix the behavior for many models. Only providing the computed relation (explicitly stating the date has passed) reliably eliminated stale framing.
Figure 2: Structure of a session-rekindling episode re-engaged on June 5, comparing retained records, unobserved event outcomes, and divergent responses from two model routes under identical clock readings.

Key Results

  • Leaderboard Performance: DeepSeek-V4.1-Flash and hy3 were the top performers, showing strong adaptation with low stale framing (0% and 19.2% respectively).
  • The Failure of Older/Smaller Models: DeepSeek-V4-Flash and Qwen3.5-397B suffered from massive dissociation, with over 84% stale framing rates despite nearly 100% accuracy in recognizing the date had passed in their reasoning traces.
  • Clock Usage: In the "Natural" condition, reasoning-enabled agents were quite good at recognizing they needed the time, invoking the Clock tool in 93.5% of cases. The bottleneck isn't getting the time; it's using it.
  • Hiatus Sensitivity: Surprisingly, a longer hiatus (120 days vs. 30 days) actually made some models more prone to reviving stale plans, suggesting that the passage of time doesn't naturally "decay" the relevance of a memory for these agents.

Limitations and Future Work

  • Synthetic/Chinese Focus: The benchmark is currently limited to synthetic Chinese service scenarios. Expanding to multi-lingual and real-world messy data is a logical next step.
  • Outcome Uncertainty: The paper assumes the outcome of the past event is "unknown." Future work could explore how agents handle explicit vs. implicit outcome updates (e.g., "I missed my appointment").
  • Black-box Constraints: Because the study used commercial APIs, the authors couldn't inspect internal activations to see why the reasoning-action gap exists at a mechanistic level.
  • Semantic Entailment: The current lexical flags for "completion" are a proxy; more robust semantic checks for how agents handle "expired" intentions are needed.

Summary updated

Abstract

Long-horizon agents increasingly serve as persistent personal assistants across modern web platforms, managing schedules, applications, and multi-step plans over weeks or months. Users often return to an ongoing conversation after a long conversational hiatus, while earlier bookings and preparation notes stay on record. During this hiatus, scheduled dates may pass even though the retained records remain unchanged and event outcomes go unrecorded. Existing evaluations of long-horizon agents focus on fact retention and explicit state updates. Whether agents perceive the passage of time and adapt their advice accordingly therefore remains unknown. We frame this ability as temporal state perception and use, which means perceiving the current time and applying it when interpreting dated records. To study it, we design session-rekindling tasks that keep the earlier conversation fixed and vary only the follow-up date and the way temporal information reaches the agent. Experiments across five model routes and five temporal-information conditions show that agents can obtain the current time but often fail to use it. They reliably query the clock and even note in their reasoning that the scheduled date has passed, yet many still tell users to prepare for the expired event. Supplying explicit temporal context reduces this error, and providing the computed date relation removes it entirely. Keeping up with time therefore depends on turning perceived time into updated plans. Based on these findings, we design a session-rekindling benchmark that provides temporal-awareness guidance for designing long-horizon agents. The code can be found at https://github.com/Adoreu2/session-rekindling-benchmark.
BibTeX
@misc{su2026can,
  author = {Hongguan Su and Ming Wang and Shun Zhang and Yifan Wang and Rong Yin and Zhiming Ma and Mengzheng Yang and Peng Chen},
  title = {{Can Agents Keep Up with Time? An Empirical Analysis of Temporal Adaptation After a Conversational Hiatus}},
  year = {2026},
  eprint = {2611.17806},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.17806},
}
Figure