LifeOS gets better by running. Every piece of work leaves a record of what was asked, what was done, what went wrong and what you thought of it. That record feeds back into the system, so the same mistake is harder to make next time.
Fix it where it lives
When something goes wrong, the fix goes where the behavior actually comes from. A preference goes into the rules the assistant loads every session. Something that must never happen goes into a hook that blocks it in code. Domain knowledge goes into the skill that does that work. A lesson written only in a notes file is forgotten once it scrolls out of view, so the fix always lands somewhere the system reads every time.
One rule holds without exception: a gate is never weakened to make a failure go away. If a check blocks something, either the work gets fixed or the check gets corrected on its merits.
How it learns
Memory captures sessions, decisions and corrections as they happen. Learning looks across many runs for patterns: which kinds of task keep going wrong, which instructions keep getting ignored, what you keep having to repeat. Each pattern becomes a proposed change to the system, and the change lands in the file that controls the behavior.
In practice
Say the assistant claims a site is live when only the upload succeeded. After it happens, the system adds a rule and a verification step: a live claim now needs two kinds of evidence, a server probe and a real-browser screenshot. From then on, a report without both says “deploy accepted” instead of “live.”