Puzzle Memory: A Versioned Task State Architecture for Long-Horizon Human-AI Collaboration

Long conversations with an AI can preserve every message and still lose the actual state of the work. A transcript records what was said, but it does not clearly identify which goal is active, which decisions have been replaced, which constraints still apply, or which questions remain open. A short summary is cheaper to use, but it can remove evidence needed to understand or correct the current plan. Puzzle Memory is a research proposal for separating three functions that are often mixed together. Raw Memory preserves historical evidence. Puzzle Memory maintains a selective and versioned representation of the current task. Working Memory contains the operational state needed for the next action. A readable Puzzle Story allows a person to inspect and correct the system's interpretation. Historical evidence may become dormant once its relevant contribution has been represented in the current task state. Dormant evidence is not deleted: it remains traceable and can be retrieved again when a contradiction appears, confidence falls, the goal changes, an earlier branch is resumed, or a person requests verification. This could reduce routine context cost while preserving a route back to the original evidence. The proposal does not claim that retrieval, summarization, task-state representations, provenance, versioning, or rollback are individually new. Its falsifiable hypothesis concerns how these mechanisms are organized around an evolving task state. The paper defines the architecture, its invariants and failure modes, and an evaluation program comparing it with strong alternative memory systems. It reports no empirical performance result at this stage.