problem.
the obstacle or failure mode. standalone. does not link to anything. future agents match against it when something similar surfaces.
procedural memory
case-based reasoning.
procedural memory records what broke, what fixed it, and what failed to fix it. three memory kinds—problems, solutions, and failed tactics—form structured cases that future agents can match against, evaluate, and apply. the next similar problem starts with experience, not a blank slate.
three kinds.
solution.
what worked. must link back to the problem it answers. without that link, the write is rejected.
failed tactic.
what was tried and did not work. also must link back to its problem. knowing what to avoid is half the solution.
the scenario.
enforced structure.
every solution and failed tactic creates a problem-attempt record linking it to the problem. this is not optional. the pipeline adds the record atomically.
utility is per-problem.
the same memory can be critical for one problem and useless for another. utility votes are recorded against specific problems. global utility is derived from these observations, never assumed.
the unit of reuse.
a scenario is a problem and its linked attempts. not isolated tips. not loose advice. a structured case the agent can read, evaluate, and apply.
accumulated experience.
procedural memory is what makes shellbrain feel like experience. the next attempt doesn't start blind. it starts with every relevant failure and every relevant fix already surfaced.