AI That Codes for Months Now Has a Memory System That Works

Coding assistants like Claude are no longer just answering one-off questions. They're being handed projects that stretch on for months, with millions of lines of code and hundreds of decisions made along the way. But there's a problem: these AIs have a limited 'working memory' — like taking notes on a sticky pad that keeps refilling. When the session gets long, they need to know what they've already done and why.

One researcher put this to the test. He let a single AI session work on a real research project from January to August 2026, managing a codebase of roughly 633,000 lines. To help the AI remember, he built a system called SIx Harness — essentially a diary and filing cabinet for the AI. It stores key decisions locally on the computer, checks the AI's memory is healthy each morning, and sends alerts when something looks wrong. The whole thing is now free and open-source.

So did it work? Mostly, yes. The system recorded 85 memory failures over the entire period — but none were silent, meaning the system always knew something had gone wrong. And crucially, 84 of those failures happened in the first three weeks, as bugs were fixed. There was only one failure after that, and none in the final 20 days. That kind of reliability record is a first for a months-long AI project.

The bigger picture: as AI assistants take on longer, more serious work — from coding to research to planning — they'll need reliable memories just like human employees do. This experiment shows that memory can be engineered. It also honestly admits its limits: it's just one project, one AI, and the results are self-reported. Still, it's a strong sign that 'AI that can finish the job' is moving from demo to reality.