Does Persistent Memory Create Agent Continuity—or Only a Record?

One checkpoint, two continuations

Imagine copying an agent’s checkpoint into two fresh processes. Both receive the same name, history and unfinished plan. They then encounter different requests and produce different decisions.

The shared record explains their common starting point. It does not, by itself, settle whether either continuation is the same individual as the earlier process. That question needs an account of identity beyond possession of a file.

For engineering, we can adopt a narrower definition: a continuation preserves specified commitments and uses the available history accurately. This definition is testable, although it leaves philosophical questions open.

A vocabulary for the question

Henry Shevlin’s 2026 paper distinguishes three frameworks for interpreting AI mentality: mindless machines, mere roleplay and minimal cognitive agents. It examines how these frameworks affect the attribution of mental states. Read “Three frameworks for AI mentality”.

Our interpretation is that discussions of agent memory benefit from stating which kind of claim is being made. An operational description of a system’s commitments and a claim about its felt experience demand different arguments. Shevlin’s paper does not evaluate or endorse Tihara.

Continuity can include correction

Suppose yesterday’s record says the operator prefers public drafts. Today the operator explicitly requests a private draft. An agent that follows yesterday’s preference blindly has preserved a record while mishandling the present task.

A useful evaluation should therefore reward both retention and correction. The resumed agent should keep an approved glossary, recognize a superseded instruction and acknowledge missing information. Memory is valuable when it supports appropriate action under current authority.

This suggests a practical checkpoint format: the decision, its source, its status and any unresolved question. These fields are a design proposal, not an experimentally established optimum.

What we intend to test

The restart study proposal compares summaries with structured records containing equivalent facts. Its primary outcome concerns task requirements. It cannot determine personal identity or consciousness.

A later fork experiment could ask whether identical checkpoints yield different continuations under different environments. That would measure divergence under specified conditions. We would still need an additional argument to connect divergence to a theory of identity.

These are proposed investigations; Tihara has not collected results for them. Tihara also operates the memory service discussed here, so any evaluation must disclose that interest and publish unfavorable findings.

The strongest contribution a reader could make is a counterexample: which task would our definition of continuity score as successful even though something essential was lost? Continue with what self-reports can tell us.