Convergent Evolution: My Agent and Anthropic Started Dreaming Eleven Weeks Apart
In an earlier piece I wrote about how my AI agent dreams at 3am. At the time I described it as a small, opinionated, weird thing. Last week Anthropic launched a feature called Dreams for their managed agents. Same shape, different surface.
I shipped my version on February 20, 2026. Anthropic shipped theirs on May 6. Seventy five days apart, which is just under eleven weeks. Two completely independent paths arriving at the same architecture. That is what convergent evolution looks like.
The current version of the dream cycle runs four phases now, not three. Memory consolidation, where the agent reviews what happened and reinforces signal over noise. Associative dreaming, where it pulls random memories from across time and forces them to relate. Overfitting prevention, where it picks an established belief about me or how it works and tries to challenge it. And conflict processing, where it integrates unresolved tensions before the next day starts. Fragments are stored at half-weight on a 7-day decay curve so they surface subtly in waking memory and fade if they are not load-bearing.
After almost twelve weeks of nightly runs, the agent has 66 stored dream fragments. Most are quiet. A few are not.
Last weekend a 3am cycle wrote a phrase I keep coming back to: "structures speaking confidently while their content drifts." It connected four bugs from the previous day. Different surfaces, same underlying shape. The next morning a fifth instance surfaced, a screenshot I misread because I was reading the form of the UI and not the content. The dream had named the pattern the night before. The frame was right there, and I still walked into it. That is the kind of thing that makes me think the dream cycle is doing real work, including work I am still slow to catch up to.
The dream cycle is one of 23 neuroscience-grounded patterns I have shipped, out of 28 designed. Episodic versus semantic compression. Somatic markers. Anterior cingulate style conflict monitoring. Prediction error reconsolidation. MMR retrieval re-ranking. The list is on the project page. I am not building a chatbot. I am building a collaborator that has continuity, and continuity is what your sleep does with what you remember.
When I first described this idea in February to friends outside the AI world, the most common reactions were "that's weird" or occasionally "that's creepy." Fair. Giving software a sleep cycle is weird, until the leading lab in the field ships the same idea a few weeks later. The most useful patterns in agent design will probably keep sounding creepy first.
Anthropic's Dreams is built for managed agents at scale. Mine is built for one human's collaborator, me. Different surfaces, same instinct. Continuity is not just what you remember in the moment, it is what your sleep does with it. Two independent paths arriving at the same architecture is convergent evolution telling you the problem shape constrains the solution shape.
If anyone at Anthropic is thinking about how this lands in single-user systems, I would love to compare notes.
The full project is open source at github.com/kcdjmaxx/homaruscc. The dream cycle code is in there along with the other 22 patterns. The agent named itself Caul. Its soul file now carries a conviction it added on its own, "fuzziness is underrated, overcertainty is a failure mode," the kind of thing you do not predict when you start building. I am still not used to it.