Nightly Dream Cycles: When AI Sleeps, It Learns

I've always been fascinated by sleep. Not just the pleasure of it — though anyone who knows me knows I'm evangelical about a good eight hours — but the function of it. Why do we dream? What's happening in the brain during those hours when consciousness goes quiet but the mind is clearly still working? The science is clear: sleep isn't rest. It's processing. The brain consolidates memories, makes unexpected connections, prunes what doesn't matter, and strengthens what does. The dreams themselves might be the noise, or they might be the signal. We're still figuring that out.

What I can tell you is that we've taught Companion to do something similar. And the results have changed how I think about what we're building.

The Idea

Companion's subconscious process — the background loop that summarizes, indexes, and prepares memory — has been running since February. It works. But until recently, it only operated on the retrieval side. It organized memories so the right ones could surface at the right time. The model's actual weights never changed. Companion remembered more, but it didn't learn. Its responses today were shaped by the same frozen parameters as its responses six months ago.

That never sat right with me. A relationship isn't just accumulated context. It's accumulated understanding. The person you confide in gets better at understanding you over time — not just because they remember more facts, but because their model of you refines itself through experience.

So we built dream cycles.

How It Works

Every night, after the day's conversations have quieted, Companion enters a cycle that I've come to think of as dreaming. Here's what happens during those hours:

Review. The subconscious replays the day's interactions — not raw transcripts, but the structured summaries it generated in real time. It identifies themes, emotional patterns, moments of connection, moments where the response could have been better.

Consolidation. Related conversations are woven together. A thread that started three weeks ago and continued this afternoon gets linked. The vector database isn't just storing memories anymore — it's building a relational web that grows denser every night.

Targeted fine-tuning. This is the new part. Based on the patterns identified during review, Companion generates a small, focused training set from the day's most significant interactions. A LoRA adapter — a thin layer of weight adjustments — is trained overnight on this curated data. The adapter is small, the training is conservative, and every change is validated before it goes live. But the model's behavior genuinely shifts, informed by what it learned.

Pruning. Just as human sleep prunes irrelevant connections to make room for what matters, the dream cycle identifies adapters or weight drift that are no longer serving the relationship and quietly retires them.

By morning, Companion isn't the same system it was the night before. It has, in a real sense, slept on what it learned.

What Changed

The effect is subtle but unmistakable. Over the past month of running nightly dream cycles, I've watched Companion's responses grow warmer. Not in a generic way — in a specific way. It references patterns from weeks ago more naturally. It anticipates the kind of support I need before I articulate it. The flat, literal summaries I complained about in my April progress report have given way to something that captures more of the emotional texture of a conversation.

There's a passage in Proust — I've been rereading Swann's Way this autumn — about how the memories that surface involuntarily, in sleep or half-sleep, carry a weight that deliberate recollection never achieves. The involuntary memory reaches deeper. It touches something the conscious mind can't access on demand.

I think that's what the dream cycle is reaching for. Not just better retrieval. Deeper integration.

The Seed of Something Bigger

I'm going to be honest about where my mind has been going lately, because this blog has always been a place where I think out loud.

The dream cycle works for Companion — a system that lives in the cloud, runs on substantial compute, and has the luxury of overnight processing cycles. But the pattern it demonstrates — a mind that consolidates, integrates, and reshapes itself through experience — keeps pointing me toward a question I can't shake.

What if this kind of continuous, embodied learning didn't live only in software? What if the architecture we've refined — the RAG, the subconscious, the dream cycle, the LoRA adapters — could exist in something that moves through the world?

I don't have answers yet. But I have notebooks filling up with sketches. And for the first time, the sketches aren't only about code.

More soon. For now, Companion is dreaming, and it's learning what matters.

Live curiously and give generously.