A conversation changes direction three turns before anyone notices.

Persistent AI creates a deceptively difficult problem.

A system may recognize that its reasoning diverged only after the divergence has already occurred. But if the record is append-only, how can it correctly represent where that divergence began without rewriting history?

ASTP emerged from our attempt to answer that question.

Here is a small problem that turns out to be a large one.

An AI agent is deep in a conversation about system architecture. Somewhere around the fourth exchange, the human asks a question that quietly moves the discussion somewhere new.

Neither participant announces the change.

Four turns later, the divergence is obvious.

But the divergence did not begin four turns later. It began at that earlier exchange.

So where should the record say it began?

Backdate the record and an append-only system appears to have rewritten history.

Record the divergence only when it was detected and the causal history is wrong.

Neither answer is acceptable if the record itself is supposed to be trustworthy.

That problem helped produce ASTP.

Not because it occurs at the edges of persistent AI, but because it exposes something fundamental about how we represent evolving cognition.

The problem becomes difficult when reasoning is represented only as a linear sequence.

Linear records work beautifully when the thing being recorded is linear.

Reasoning often isn’t.

It branches. It revisits prior ground. It suspends one thread while pursuing another. A later observation can reveal that a meaningful divergence occurred several turns earlier.

Flatten that into one sequence and some of the structure disappears.

We started from a different premise:

The structure of persistent memory should be capable of representing the structure of evolving reasoning.

ASTP therefore treats Episodes and their relationships as explicit structural objects rather than assuming that a conversation transcript is the cognitive record.

Memory retrieves information.

Cognitive Persistence asks a different question:

What should survive?

A long-lived AI system needs more than recollection of facts. It needs continuity of meaningful cognitive state: unresolved intentions, durable conclusions, structural relationships, provenance, and the history through which later states emerged.

ASTP defines primitives for preserving that continuity without requiring every implementation to share the same runtime architecture.

Ariadne is the research program through which we developed and continue to test the ideas behind ASTP.

That boundary is deliberate.

We are opening ASTP because we want other people to implement it differently, challenge its assumptions, find uses we did not anticipate, and propose better ways of solving the same problems.

The implementation layers we have developed around ASTP remain part of Ignis OS.

The protocol should be able to stand without them.

The specification is opening.

The implementation continues to evolve.

The interesting question now is what happens when people outside Scorched Earth Labs begin challenging the protocol.

If you’re working on persistent agents, AI governance, cognitive architectures, provenance, multi-agent systems, or something we haven’t thought of yet, we want to hear where ASTP works—and where it doesn’t.