Why I taught my agents to dream

I didn't set out to make my agents dream. I set out to stop them from lying to me.

I run a second brain. A vault of plain markdown files that my AI agents treat as source memory. Threat intel, competitive landscape, field stories, positions I've taken in front of rooms. Every night, automated scouts sweep for new findings and stage them. Every week, the best of it gets promoted into long-term memory. The pipeline hums along whether I'm awake or not.

For a while I thought that was the whole job. Collect new intel, compress it, promote the good stuff into memory. More memory, smarter agents.

Then I caught one of my agents briefing me on a threat in the future tense. Attackers will eventually be able to do this, it told me, confident and well-sourced. Except it had already happened, and the proof was sitting in the same vault, staged by my own scouts days earlier. The agent's core beliefs had simply never been told.

That's the failure mode nobody warns you about. An agent with stale memory doesn't fail loudly. It briefs you with conviction about a world that no longer exists. In my line of work, the gap between "researchers demonstrated" and "actively exploited" can be a single news cycle, so that conviction is the dangerous part.

The reverse question

Everyone building agent memory right now is solving for ingestion. How do we get more context in, faster, with better retrieval. Almost nobody is asking the reverse question: does anything we learned this week contradict what the agent already believes?

Your brain asks that question every night. Sleep researchers have been saying for decades that dreaming, whatever else it is, is partly memory maintenance. Replay, consolidation, reconciling the new against the old. You go to bed with contradictions and wake up with fewer of them.

My agents had no equivalent. They had perfect recall and zero reflection. So I built the missing pass, and the name stuck before I could come up with anything more professional. The dream pass.

The dream pass

Here's what it actually does. Every Saturday morning, before I'm useful to anyone, each agent reads its own core beliefs. Not the intel, its beliefs: the file that defines what it holds true about threats, competitors, and doctrine. Then it audits those beliefs against the freshest intel corpus and writes me a ranked gap report.

The gaps fall into four buckets. Direct contradiction is the five-alarm one, where the core conflicts with what is now true. Missing class means a threat or capability the agent is blind to. Then there are stale facts, numbers and examples that got overtaken, and competitive drift. Every gap gets a severity, an effort estimate, and a citation, and the citation rule is non-negotiable. Without it the agent starts reporting opinions as findings, and now you have a different problem.

For anything critical and cheap to fix, it drafts the exact patch, old text and new text side by side, so approving takes seconds.

And then it stops. That's the part I had to build on purpose, because nothing about these systems stops on its own.

The dream pass never edits its own memory, not one line. Every proposed patch sits in a report marked pending until I walk it, gap by gap, and approve, skip, or edit.

And the judgment I supply isn't only approve or skip on the wording. It's whether the intel that triggered the gap can be trusted at all. A scout can stage something thin or wrong, and a contradiction built on a bad source still gets flagged with full confidence. The pass surfaces the conflict. It doesn't vouch for the corpus.

That gate is not me being precious about my markdown. It's the whole design. The first time this pass ran, it found nine gaps in one agent's beliefs, and working the list surfaced two more. By the end of the cycle we had closed eleven, and the agent's core grew from 221 lines to 263. Every one of those lines passed through a human, which means I can tell you what changed, when, why, and what evidence drove it. The memory has a chain of custody.

What a gap looks like

A recent pass caught my threat-analyst agent holding the belief that AI-assisted exploit development was a frontier research capability. Real and demonstrated, but a lab story. The week's intel said otherwise: a documented, production-grade operation building modular payloads with commodity AI tooling. The dream pass flagged it as a tense problem. "Will be able to" had become "did." If you present to boards, you know that tense change is the entire difference between a future-risk slide and an incident-readiness conversation.

Here is that entry the way it landed in the report, sanitized only where it named my files:

I approved that patch in about a minute. That's why severity gets paired with effort. The agent did the forensic work overnight, and I supplied the judgment over coffee.

The trade

Everyone building agents right now wants them to think for themselves. Tell a model to ponder overnight, leave it running while you sleep, and see what it comes back with. People I trust have watched that produce something the next morning that's part brilliant and part unhinged. That's the trade as I see it. Unstructured reflection produces magic you can't reproduce. Governed reflection produces fewer fireworks, and it shows up every Saturday with receipts.

I'd rather have the boring version. Not because the magic isn't real, but because I have to stand behind what these agents tell me. When an agent's output ends up in front of a CISO, "it emerged overnight and I kept the good parts" is not a provenance story. A ranked gap report with citations and an approval trail is.

If your agents have memory, they have drift. The only question is whether you've built the thing that catches it, and whether that thing answers to you. Reflection without a gate is just drift with good PR.

You find out your agent's memory is stale either before it briefs the room or after. Only one of those is recoverable.

A note on how this was written

Fair warning, because some of you will clock it anyway: the first draft of this piece came from the same agent system it describes, working from a voice file built on transcripts of how I actually talk. A friend read an early version and told me he has grown tired of punchy AI writing. He had a point, and it's the same point the piece makes. So the draft went through the gate like everything else. A human walked it line by line, cut what didn't sound like me, and signed off on what did. You're reading a governance essay that practiced what it preached, and I'll take that trade too.