[MODEL] Weighting of memory/docs, and maybe their sources, too heavy compared to "checking it live"

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude made incorrect assumptions about my project

What You Asked Claude to Do

Remember things and also be heavy on "trust but verify" (in many different formats)

What Claude Actually Did

Seems to honor memory or docs way more than live checks.

If it had a thermometer / sensor outside and last checked the temp in Jan, and I asked what temp it is, I feel like it would just latch onto the logged -10f instead of the current April ~50f that is really no more effort than checking docs.

Expected Behavior

I feel that some weighting is in place, and it "feels" more heavy lately, to where it trusts docs and memories without any proper sense of time or volatility. I see this with lots of things from state of an environment, number of files in a directory, major architecture decisions, what host it is running on, and more.

I am rarely one to say "it changed" because I both know it does and this is a great time to be alive with the rapid evolution of not only the model but everything around it, and that I am also "everything around it". If I had to guess, I would guess that the auto memory and dream feature testing/releases were about when I "felt" the change and now it's definitely annoying. I'm really avoiding defensive logic in hooks and triggers hoping it gets smoothed out in the right way, and still trying to find the best per-project CLAUDE.md or other injections to keep the dude grounded to reality.

Ranting a little more; it feels like the lack of real sense of time is a major part of it. I started to demo a rigging to token-cheaply force claude to know the time passed since each action and prompt and stuff but found it wasn't really feeling right, because those numbers are seemingly meaningless.

Anywho, if yall read this one in the firehose, thanks for the read if nothing else! :)

Claude Model

Opus

Platform

Anthropic API