Most systems remember failure too aggressively. An action fails once, a branch scores poorly, an agent cannot complete a task, or a strategy becomes infeasible. The system compresses that event into a blunt conclusion: do not try this again.
Conditional Failure Memory keeps the useful part of the failure without turning yesterday’s conditions into tomorrow’s law.
The minimum record
Instead of storing only failed, the system records the possibility, the state in which it was attempted, the relevant constraints, the observed failure mode, and the evidence supporting that diagnosis. The object being remembered is therefore not “X does not work.” It is closer to “X did not work under conditions C for reason R, with evidence E.”
That distinction changes what memory is allowed to do. Failure can lower priority while the blocking conditions remain true without deleting the branch from the recognized possibility space.
Graded dormancy
A temporarily bad possibility does not need to compete for full attention. It can become dormant. Dormancy is graded rather than binary: a branch can be deeply inactive when its blockers are stable, lightly dormant when conditions are close to changing, or active when current evidence makes another attempt worthwhile.
This is the companion principle expressed publicly as DORMANT ≠ DEAD. Memory suppresses wasted repetition while preserving the option.
Causal reactivation
The mechanism becomes useful when state changes can wake affected possibilities. If cost falls, a dependency becomes available, a capability improves, a deadline moves, new evidence arrives, or another branch changes the environment, the system asks which prior failures depended on that condition.
Only causally relevant changes should reactivate a branch. A random change somewhere in the system is not enough. The remembered failure conditions provide the bridge between new state and reconsideration.
Why this matters for agents and decision systems
An agent that permanently blacklists failed approaches becomes brittle. An agent that forgets every failure loops. Conditional Failure Memory creates a third behavior: retain the lesson, preserve the possibility, and spend new compute when the world has changed enough to justify it.
That same structure applies beyond software agents. A project, route, collaboration, funding strategy, creative direction, or personal action can be unavailable now without being intrinsically impossible. The decision-relevant question is what made it fail and whether those governing conditions still hold.
Its place in Possibility Reserve
Conditional Failure Memory is one part of a larger architecture. Graded Dormancy controls how inactive possibilities are retained. Causal Reactivation determines when changed conditions should wake them. Path Expansion gives additional value to branches that unlock desirable future states, even when they are not the immediate answer.
Together, those mechanisms let a system learn from failure without allowing failure to permanently collapse its future.