Why does my coding agent keep making the same mistakes?

Because nothing tells it the mistake was a mistake. A fresh session cannot see last week's failed attempt, and even saved notes carry no signal for "this was tried and it broke." Memory has to record outcomes, not just decisions — a claim that failed must lose standing, and the trap must be served before the agent walks into it again.

the missing signal failure is exactly what nobody pauses to document

When a fix breaks in review, the knowledge "that approach fails here" usually lives in your head and nowhere else. The next session — or a teammate's agent — meets the same code with the same training and the same blind spot, and produces the same confident mistake. Notes help only if someone wrote the failure down, and nobody documents failure mid-failure.

decisions are not enough the record has to hear back

A memory of decisions still repeats mistakes: the decision that turned out wrong is remembered as confidently as the one that held. In memcell, every served statement is judged against what the session actually did with it — no one files a review by hand. A rule the session ran on draws worked (+0.15); a claim the work showed to be wrong draws failed (−0.25), with the arithmetic on the chain; a trap the session steered around draws avoided.

"Retry the webhook by re-posting the event."      0.70
session 41 followed it and lost the event         −0.25
                                                   0.45
below the space's floor → readable, not served

drift a rule ignored is a rule said again, not a rule doubted

The other repeat is an agent going against a rule it was already given. That is not evidence the rule is wrong, so it moves no number. It is recorded as a divergence against the session, and the statement returns on that session's next recall, marked and leading what memory serves — whatever the next question happens to be.

A memory that dropped confidence here would punish a rule for being disobeyed, and the most-flouted rule would be the first to fall silent.

retirement by evidence nobody has to remember to clean up

Confidence is the memory's honesty about itself: a statement that keeps failing falls under the serve floor and stops being repeated — retired by evidence, not by someone noticing. Measured from an empty space on the live loop, rule violations halved after one turn and held (0.50 → 0.25 → 0.25); with the loop severed they climbed instead (0.75 → 1.00 → 1.00).

npx memcell connect