Workbench / Issues

AI Forgets Earlier Constraints

A constraint, instruction, preference, or decision that should persist through the task stops affecting later output.

What This Looks Like

The user gives a constraint, instruction, preference, format rule, exclusion, or decision that should continue to govern the task, but later output stops following it. The AI may initially comply and then drift, ignore an earlier limit, reintroduce excluded content, or behave as if a prior decision was never made.

Why It Matters

Many AI workflows depend on constraints persisting across multiple turns, steps, files, or generated artifacts. When earlier constraints stop applying, users have to repeat themselves, audit prior decisions, and check whether later output is still governed by the task contract.

Structural Signal

A constraint was declared earlier and should remain active, but it loses governing effect in a later state. The issue is not simply that the AI made one mistake; it is that an active constraint failed to persist across the task.

Common Triggers

When to Use This Issue

Use this Issue when a prior constraint should still govern the work but stops affecting later output, behavior, formatting, or decisions.

When Not to Use This Issue

Do not use this Issue when the user changed the instruction, intentionally relaxed the constraint, or never clearly declared the requirement. Do not use it for ordinary context loss unless a specific earlier constraint stopped applying.

Category

Memory & Context

Primary Pattern

PAT-0350 — Persistence Instability

Declared Patterns

Derived Primary Lenses

Derived Secondary Lenses

Search Intents

Ontology Metadata

Code
ISS-0029
Version
ISS-0029@0.1.0
Ontology release
0.1.0
Updated
2026-05-10T00:00:00Z

History

No public history entries recorded.

View full ontology changelog