Summary / The 27 components
20 Episodic coding memory
Remembering prior attempts, mistakes and successes across tasks.
Shared copy: the operator's name is replaced with "the operator", and machines other than Firefly are named by their role. Nothing else is changed.
What is built now
A CodingEpisode row is written at MEMORIZE for every harness task (80 rows).
What is missing
Lesson distillation, scope-matched retrieval, helped/hurt bookkeeping, and proposing lessons into her main memory through its gate.
Reviewer Claude, 2026-10-04
Reviewer's read: Episodes recorded; nothing reads them
Writing episodes cost nothing and keeps the option open. Reading them back is where the risk lies: injected lessons are a classic way to make a benchmark look better through contamination, because SWE-bench tasks in one repository share code. Any memory arm must keep held-out tasks out of the store.
How it is measured recording only
Memory arm against h1-v41, with the store frozen before the held-out set runs.
Adoption rules shared by every arm: How the tests work.
Research & attempts
Score history
No scores yet. The first entry is the baseline for suite v1.
Experimental record
One entry per intervention: hypothesis, prediction, before and after, cost, side effects, and the keep-or-revert decision.
No interventions recorded yet.