Continuity

A stateless chatbot is amnesiac by design. Every conversation begins with nothing: no memory of past decisions, no record of what was tried, no way to distinguish “nothing exists” from “the system failed to look.” That amnesia is fine for casual chat. It is disqualifying for anything that acts in the real world.

Businesses run on continuity. A missed follow-up, a forgotten commitment, a decision whose reasoning evaporated: these are continuity failures, and they cost real money. An AI system that serves a business must have the same property the business needs: it must remember.

Continuity in PING is not a single feature. It is a stack of commitments:

• Events are source truth. Things that happen are recorded as events with identity, timestamp, and provenance. The event log is the system's memory of what occurred.
• Knowledge accumulates. Events are distilled into canonical knowledge: observations, claims, relationships. New evidence updates knowledge; contradictions are preserved, not silently overwritten.
• Evidence is attached. Every derived claim links back to the source it came from. You can always ask “where did this come from?” and get an answer.
• State is replayable. Given the same causal history, the system should reach the same canonical state. Replay is how continuity is verified, not just asserted.

“Memory” in AI products usually means a longer context window or a vector store of past chats. That is retrieval, not continuity. Continuity adds the properties retrieval lacks: identity across transformations, explicit provenance, typed failure states (so “not found” never hides “lookup failed”), and deterministic replay.

The test is simple: can the system explain, after the fact, what happened, why it happened, which version of the system decided, and show the evidence? If yes, it has continuity. If it can only show you a chat log, it has retrieval.

Continuity is the oldest and most documented concept in PING. The event pipeline, canonical knowledge layer, and replay semantics are designed around it. The blog publishes implementation notes as the machinery matures.