Ideas

The interesting problems are mostly not chat boxes.

Long-running AI work needs infrastructure for identity, memory, decisions, evidence and recovery.

Every meaningful output should have a stable identity, checksum, lineage and relationship to the conversation that produced it.

Approved, rejected, superseded and conflicting choices should form an inspectable graph rather than disappear into chronology.

Generate the smallest complete continuation package for a specific task instead of dumping an entire history into every session.

Claims should carry the difference between sourced fact, model knowledge, inference and uncertainty.

Irreversible or high-impact actions deserve explicit human review tied to the exact proposed change.

Organizations need to retain not only incidents and documents but why they responded as they did.