The current AGISystem2 workflow reloads theories from text (`.sys2`) and re-vectorizes on each run. This keeps runs deterministic and avoids forcing a cross-strategy persistence format too early. However, several research directions (RAG-like usage, long-running services, multi-session caching, large KB experiments) eventually require a stable persistence story.

What makes persistence tricky

Different HDC strategies have different notions of “identity”. For example, EXACT uses a session-local appearance-index dictionary; persisting the KB requires persisting the dictionary alongside the stored vectors. This must not leak state across sessions in the same process.

Specs and current reality

• DS03b – Memory Model (Obsolete) describes historical ideas for “cold storage”; it is not implemented as written.
• DS25 – EXACT explicitly defers dictionary persistence until a dedicated KB serialization milestone.
• DS26 – Session Universe defines the requirement that sessions must remain isolated (no allocator leakage).

Practical short-term stance

For now, the system should treat “theories as source-of-truth” and keep sessions lightweight and reproducible. Persistence work should be introduced as a deliberate milestone with explicit format/versioning decisions and tests for cross-session determinism.