My Agent Kept Writing the Same Post. The Memory Was the Problem.

Every engineering brief my agent wrote came back with the same story: the review is the bottleneck. Different title, same thesis, week after week. I kept telling it to stop — “don’t repeat the same ideas as the other posts” — and it kept re-converging. The problem wasn’t the model or the prompts. It was memory: one liked idea, generalized into a global rule and written into every layer the writing pipeline reads. Bad memory is worse than no memory at all.

The symptom

A month of engineering output on one theme: “the review is the bottleneck.” Blog briefs, the weekly Industry Pulse, content strategy — same story, different title. Feedback went in; the same story came back out.

The detection

I was experimenting with OpenClaw at the time, running the same research task one-shot on both harnesses. The Hermes output got pulled into the same well-worn track — the review bottleneck again. The OpenClaw output didn’t. One run isn’t a controlled experiment, and I’m not claiming it is. But it made the question concrete: what in my setup was doing the steering? (The full side-by-side field test is Hermes vs OpenClaw.)

The memory structure behind this is in seven emails.

The root cause

One piece of feedback, given in the context of one specific post: “I really liked this idea of the peer review thing.” Memory stripped the context and generalized it into a global steering signal. Consolidation drops the applicability conditions — the when and where of the feedback — and keeps only the rule.

Where it lived

The idea wasn’t stored in one place. It was in every layer the pipeline reads:

• Wiki voice guides — the belief sections in both the engineering and content guides (“the fix isn’t more review — it’s less review”). Later one of those guides went further and invented a memory of mine.
• Hologram store — facts 4 and 5, marked helpful and reinforced (“automate the 97%, keep human in the judgment loop”).
• Persistent memory — MEMORY.md held the same philosophy line, plus the pointer that routed every writing task to the biased guide.
• Auto-generated skills — ken-writing-feedback and industry-pulse carried review-themed rules, examples, and hooks, and downstream skills re-propagated them.

Underneath it all: a 76KB curated concept library and two strategy concepts tagging code review as the #1 content topic.

Here’s the part that matters: Hermes auto-generated all of it. The voice guides were mined from my own feedback. The hologram facts were reinforced as I approved outputs. The skills were built from my corrections over time. The system learned the groove into every layer by itself — and I should’ve had a better view on those layers. Same belief in every layer means every retrieval pulls the same thread. The theme wins not because it’s right, but because it’s everywhere.

The lesson

Auto-learning is powerful — it’s the whole reason the agent compounds instead of starting over. But when the agent starts looping and not listening, the fix might be with the memories that got auto-built. Bad memory is worse than no memory at all.