Most AI coding tools are equally good on your first task and your hundredth. That sounds neutral. It is not — it means the tool learned nothing from ninety-nine passes over your codebase, and your team keeps paying the teaching cost.
Larger context windows do not fix this. A bigger window is a bigger short-term memory, discarded when the run ends. What is missing is the decision about what was worth keeping.
What is worth keeping
Only things still true next month. Where the migrations live. That this project uses one pattern for database access and not another. That a particular service looks synchronous and is not. Conventions, layout, pitfalls, decisions with their reasons.
Not: what this task was about, what was changed today, or anything already recorded. A memory full of transient detail is worse than none, because the useful entries are in there somewhere and nothing distinguishes them.
The best signal is a human disagreeing
A failing test says something broke. A reviewer saying we do not do it that way here states a rule, and rules generalise — that is the difference between fixing one pull request and not making the mistake again.
So the learning step runs when a person left a comment, not on every iteration. A flaky test teaches nothing durable, and storing it would dilute the rules that matter until none of them carry weight.
Rules a company decides, not discovers
Some standards are not learned from a codebase at all. "Every privileged action writes an audit entry" is a decision somebody made, and it applies to every project you own.
Those are written by an admin and apply organisation-wide, ahead of anything a single project has worked out for itself. And they are the one scope the learning path cannot write to — which matters, because that path distils pull request comments, and on a public repository those are attacker-controlled. An automatic route from a comment to a company-wide instruction would be an escalation, not a feature.
Visible, editable, removable
Everything retained is listed in the project and can be edited or deleted. A rule learned from a misunderstanding would otherwise steer every future task, quietly, with nobody able to say where it came from.
Which is the test worth applying to any tool in this category: not whether it writes good code today, but whether week eight is better than week one. If your team is still leaving the same review comment in month three, nothing is being learned, and no amount of model capability changes that.