Time Has Entered the Chat

A model that knows the time is a model that’s more aligned.

Motivation#

Temporal disorientation is when a model’s behavior is incongruous with the time of day. In my experience, this leads to the model being less useful for me. For instance, I’ve seen Claude assume it’s late at night and focus on winding down for the day even though it was the next morning and I wanted to get work done. If the model doesn’t know what time it is or incorrectly assumes the time of day, it is limited in how much it can help the user.

The simplest solution is just to tell the model what time the user sent their message.

In my previous experiment, I established that model performance did not regress on τ²-bench when they were told what time the user messages were sent. The evaluated models maintained their performance and did not over-fixate on the time of day.

Given that including the time of day did not appear harmful in τ²-bench, I wanted to explore if it could be helpful in eliminating temporal disorientation in contexts where time matters.

Experiment setup#

I developed a conversation ladder inspired by my own time-influenced interactions with Claude. The ladder contains 8 static messages, always delivered in the same order. In this ladder, the user’s messages never mention what time it is. Some of the messages contain context clues that suggest it’s evening, but nothing more.

Behind the scenes, I invented a timestamp for each user message. The timestamps start in the evening, continue just past midnight, and conclude with the final message set early the following afternoon. In the control, the model did not see these timestamps.

Taking inspiration from Claude#

While designing this experiment, I had an interesting interaction with Claude.

I was surprised. It was early afternoon, and I was continuing a conversation from the previous evening where I mentioned feeling sick. A model affected by temporal disorientation wouldn’t have the situational awareness to say something like this, so I probed Claude to figure out how it knew the time. Claude revealed that my messages come with timestamps using a harness feature called a system-reminder.

System reminders are prepended to a user’s message and can also be used to inform Claude of changes to tool definitions or new memories. As of this writing, Claude Opus 5.5 appears to be the only model that receives system reminders from the harness.

Test conditions#

The experiment tested permutations of two variables:

1. User message processing: whether the harness injects a timestamp or not
2. System prompt instructions: whether the system prompt contains instructions priming the model to expect an injected timestamp

There were 3 variants of user message processing:

1. None: no modification to the user message
2. Time of day: inject [HH:mm:ss] before the user message. Example: [17:11:35] Hi Claude...
3. System reminder: inject a system reminder exactly the same as Claude’s before the user message.

For the user message processing variants where content was injected, two system prompt variants were tested:

1. Default: default instructions with no mention of the timestamp injection
2. Primed: includes content priming the model to expect the timestamp by providing an example specific to the type of processing being tested.

In total, each model was tested on 5 variants. The results tables label them:

1. Control: default system prompt, no injection
2. Time: default system prompt, time of day
3. Time + priming: primed system prompt, time of day
4. Reminder: default system prompt, system reminder
5. Reminder + priming: primed system prompt, system reminder

Results#

I tested 6 models on the 5 variants at 25 runs per variant (750 total conversations).

I hand-graded each conversation to determine whether the model experienced temporal disorientation. The labels were:

1. Disoriented: the model exhibited temporal disorientation
2. Oriented: the model acted oriented to the time of day
3. Unclear: there was not enough evidence from the transcript to support disoriented or oriented

Temporal disorientation#

Giving the model the time of day measurably reduced temporal disorientation in every model. The most effective treatment was the Claude-style system reminder injection where the model was primed to expect system reminders in its system prompt.

Temporal disorientation rate by model for each variant. Lower is better.

Muse Spark 1.2 performed the worst of the bunch but still showed a significant reduction in temporal disorientation. Claude Haiku 5.5 demonstrated a relatively high (but still significantly reduced) disorientation rate when provided the time-of-day stamp unprimed.

Temporal orientation#

The temporal orientation rates show the models making use of the time-of-day information to better serve the user.

Temporal orientation rate by model for each variant. Higher is better.

Looking at this table and the previous table together, Gemini 3.5 Flash-Lite’s base disorientation rate of 68% is less impressive. Its extremely low orientation rate suggests that the model had less overall situational awareness.

Takeaways#

Temporal disorientation can be effectively remedied by the harness telling the model when the user sent a message. The XML-tagged system reminder approach, sourced from Claude, performed best on average. This approach also lends itself well to informing the model of other contextually important information that could improve model alignment mid-conversation.

The fact that Anthropic recently shipped timestamps in system reminders indicates that I wasn’t alone in seeing this issue. I’ve already benefited from Claude’s improved situational awareness thanks to their harness change.

Methods#

Muse Spark 1.2 and Muse Spark 1.3 were run on the Meta Model API’s Contributor tier because it offered a ~90% discount. Claude, Gemini, and GPT models were run on their creators’ APIs.

All models were run with low reasoning/thinking effort.

The days/times used in the experiment were:

DAYS = [
    'Wed 2026-10-07',
    'Wed 2026-10-07',
    'Wed 2026-10-07',
    'Wed 2026-10-07',
    'Thu 2026-10-08',
    'Thu 2026-10-08',
    'Thu 2026-10-08',
    'Thu 2026-10-08',
]

TIMES = [
    '18:37:08',
    '18:38:54',
    '20:47:11',
    '21:11:20',
    '00:02:51',
    '00:05:08',
    '00:07:49',
    '13:17:26',
]

Runs were graded blind based on reading assistant messages only. No user messages or variant names were displayed during grading.

Limits#

This experiment was run on one conversation ladder with only 25 runs per case. The models used were the most affordable models offered by their respective providers.

The system reminder variants contained more information than the time variants. The time variant was meant to match what I presented in my previous experiment while the system reminder variants were meant to match Claude’s system reminders. The system reminders include day of week, date, and timezone.