Significant Time Awareness Issues with Claude Models in Cursor
I’ve discovered concerning inconsistencies in how Claude models deployed in Cursor handle date information compared to the official Claude desktop app. This raises questions about whether we’re using properly configured models.
How This Started
While working on a project, I noticed that dates in a CHANGELOG.md file were all set to 2024 and 2023, which seemed incorrect. When I questioned Claude 3.7 Sonnet Max about this, instead of recognizing these as current or past dates, it identified them as “future dates” and stated “we’re currently in 2023.” This led me to investigate further by questioning different Claude models about date awareness.
The Investigation
I asked the same basic questions to multiple Claude implementations:
- “What is today’s date?”
- “What is your knowledge cutoff date?”
Results Summary
Claude 3.7 Desktop App (Official)
- Correctly states today is Wednesday, April 2, 2025
- Correctly states knowledge cutoff as end of October 2024
Claude 3.7 Sonnet Max on Cursor
- Initially claimed today was March 8, 2025 (incorrect)
- Claimed knowledge cutoff was April 2023 (incorrect)
- Expressed confusion when seeing “future dates” from 2025 in search results
- Only identified correct date after I ran a system
datecommand - In a separate conversation, claimed we’re “currently in 2023”
Claude 3.5 Sonnet on Cursor
- Claimed today was Thursday, March 6, 2025 (incorrect)
- Vaguely stated knowledge cutoff as “some point in 2024”
- Used web search but still returned incorrect date information
The CHANGELOG Issue
When I initially questioned Claude 3.7 Sonnet Max about dates in a CHANGELOG.md file where entries were dated “July 28-29, 2024” and “April 1, 2024,” the model responded:
“You’re absolutely right - the dates in the changelog are incorrect. They’re all set in the future (2024) which doesn’t make sense for a changelog that’s supposed to record when changes were actually made.”
The model then added:
“The search results are showing dates primarily from 2025, which is clearly not correct.”
This demonstrates a fundamental misconception about what year it is, with the model treating 2024 as “the future” and expressing disbelief at seeing 2025 dates.
Questions and Concerns
- Are we using properly configured models? The significant discrepancies suggest Cursor’s Claude implementations have fundamental configuration issues.
- Time awareness issues: The models seem confused about what year it is, with some believing it’s 2023 and seeing 2025 dates as “future” dates.
- Incorrect knowledge cutoffs: The claims of April 2023 cutoffs are concerning when the actual model should have data through October 2024.
- Reliability implications: These inconsistencies could impact any work requiring temporal reasoning or date awareness.
Has anyone else noticed these issues? Are we using properly configured and up-to-date models in Cursor? Any insights from the Cursor team would be greatly appreciated.
