The loop that learns — a series on AI dev loops that remember, verify, and enforce, what happened when the method was turned on its own tool, what other agent setups and recent research say about keeping a setup true, what a review of the loop found it still cannot do, and where the 2026 research leaves it. Every measured result traces to a committed receipt in the public repo it describes, or to a linked source; figures from my own store or notes say so and give their date.
How one developer's AI dev loop remembers: lessons recorded once, retrieved per task, across repos, with A/B receipts. Part 1 of the loop-that-learns series.
The verifier fails too: exit codes, green builds, and judges that lie. Four rungs of independence for the verify step. Part 2 of the loop-that-learns series.
A checklist nobody is forced to run is documentation, not control: hooks, gates, and installers that wire instead of instruct. Part 3 of the series.
The method, pointed at the tool that implements it: four defects in surfaces nobody had run, and one carefully argued change that only a measurement caught. Part 4 of the series.
Notes from about 25 coding-agent tools and published setups and six recent papers: agents follow what is in front of them, skip what they have to fetch, and almost nothing checks whether a rule is still true. Part 5 of the loop-that-learns series.
A review of the loop against its own intent, and a check of that review: the read side works, the store only grows, and my summary of the review needed 32 corrections. Part 6 of the loop-that-learns series.
Seven 2026 papers on agent memory, read from the PDFs: memory alone does not force compliance, self-improving harnesses invent rules for failures that never happened, and abstraction is what transfers between codebases. Where verified memory stands beside each, the claim I had wrong, and the number okl does not yet have for itself. Part 7 of the loop-that-learns series.