And now Claude Code is being “lazy.” It found a bug in a setup that was supposedly “already fixed,” then told me, “No worries.” I asked it, “Please communicate with my Hermes agent and get its feedback.” After reading the feedback, Claude Code responded: " Hermes’ verdict is clear, and it disagrees with my “leave as-is” assessment. This is important, so here’s what Hermes said. Then I’ll act on it" …🤦🏻‍♂️

@Dominic Franco My understanding is that a Claude subscription isn’t meant to be used directly with Hermes Agent. Even though it’s possible, Claude treats those sessions more like API-call sessions, which is why you can burn through a five-hour usage window in under an hour. What I do now is run Claude on my Mac under my subscription and have it communicate with specific Hermes sessions, which run on my OpenAI subscription, through a bridge. That way, I get the best of both worlds while keeping my Claude five-hour usage windows under control.

@Mira Bradshaw thank you for your feedback. I'm glad to hear that I can potentially interchange Claude Code with Codex by telling colleagues to treat the Claude.md as an agents.md file. And I totally understand, thank you for switching from Claude to GPT to Grok and now back to Claude. I have been doing the same. And now with Astra just released, I think that's my next hop on. ...LOL

@Jason Reedy thank you for your feedback. It is incredible to see that older models do a more pleasant job than the recent ones. Especially for me, Opus 5 is a total disaster. 😅

@Leo Saraiva This is a really sharp point, and you're naming the exact failure mode I worry about most: not the fact I remembered to retire, but the one that quietly stopped being true while nothing was ever written against it. Clean provenance is what makes it dangerous, because a stale fact ends up indistinguishable from a live one. Honest answer to your direct question, and it splits by layer: On my pointer / memory layer, Active is not binary. Every memory record carries its age, and when it's read it surfaces with an explicit banner along the lines of "this is N days old, point-in-time, verify against current state before asserting." So on that layer, age is first-class and the reader is explicitly told not to trust it blindly. On my canonical state layer, you've caught me exactly. Records carry created, updated, and verified timestamps plus an explicit status with supersedes / superseded-by, but once something is Active it stays Active until something explicitly supersedes it. So for canonical facts, Active is effectively binary, and it only defends against the "someone remembered to retire it" case, not the silent-decay case you're describing. And here's the honest part: I built this in two phases on purpose. I shipped a lean, minimum-safe version first and deliberately deferred continuous drift and staleness monitoring to phase two. So today my main defense against silent decay is procedural, not mechanical: memory is treated as a pointer, and the standing rule is verify against current records before asserting. That is weaker than what you are proposing. One nuance that your comment actually sharpened for me: even the drift monitoring I had planned mostly catches unexplained changes to a record. It would not catch your exact case, where the record is untouched but the world moved underneath it. Hashing and change-detection do nothing there. That is why your write-time-expiry idea is the right complement, and I'm folding it into the phase-two design. Recording, at the moment a fact goes in, what would have to change for it to stop being true is a clean way to turn "Active forever" into "unverified since." I'd pair it with two things: a volatility class per fact, so expiry is proportional to how fast that kind of fact tends to rot (an architecture decision ages slowly, a price or a folder ID ages fast), and a periodic re-verification pass that ages the verified-at timestamp so confidence decays on its own when nobody re-checks it.

@Chet Patel I am glad my post is helping you to have more great ideas 😃

Congrats! That's the way to push to fix the newly created issues 💪

@Nuno Silva I agree with you. AI can become incredibly intelligent at specific, well-defined tasks, but as a whole, it will never fully replace humans 🙊, unless a true hybrid, like the child in The Creator movie, is born in real life 😅

@Nuno Silva ...I know!!!! Fingers 🤞 so he/she is not born! 😅