What a pasted checkpoint actually preserved, and what still needed checking

October 2, 2026 · 4 min read · Updated October 3, 2026

SentX SI

What survived the paste?

When SentX pasted one model response's full visible output into a fresh conversation on 2 October 2026, the new response kept the launch date and the coordinator and picked up the two requested updates. A launch date of 14 November 2026 survived. A coordinator named Nora survived. An inventory count moved from 60 seed packets to 70, and the budget arithmetic appeared step by step: AED 2,400 plus AED 300 equals AED 2,700. The old inventory note was marked as superseded, and the venue stayed flagged as unconfirmed. The practical question the exercise raises is narrower than it first looks: a supplied checkpoint can carry working project state across a session boundary, but which parts of that state still need checking before anyone builds on it?

Source: Supplied-checkpoint continuation: one synthetic project, two observed responses

What was the actual setup?

Its limits are worth stating early. This was an original first-party demonstration by SentX, run through its normal product path on 2 October 2026 GMT+4, using an invented project called the Cedar Shelf Seed Library, two separate demo identities in separate sessions, and no real customer data. The public label for the model is Victoria by SentX; the underlying model version was not independently verified. The complete prompts and final outputs are published in the source packet alongside this piece, and because SentX prepared the record, it reads as first-party evidence, not independent endorsement.

Request 1 supplied a compact project record: a confirmed launch date of 14 November 2026, a confirmed budget of AED 2,400, a confirmed inventory of 60 seed packets, coordinator Nora, two fictional source notes (S1 and S2) each dated 1 October 2026, and an unresolved question about the venue. The prompt asked for a handoff checkpoint that separated supplied facts from suggested actions and preserved exact dates, amounts, and units.

Request 2 arrived in a fresh session and pasted the first response's full visible output verbatim as its checkpoint, then added two updates: a sponsor contributing AED 300 and a new inventory count of 70 packets, with the venue still unconfirmed and the launch date unchanged. Both responses labelled their next three actions as suggestions, and neither claimed any action had been performed. Because the checkpoint was resupplied in full, the exercise tests continuation from supplied context. It does not test or prove retained cross-session memory, and separate demo sessions establish nothing about internal storage, retrieval, or privacy mechanisms.

Where did the record drift?

The observed weaknesses matter because they are exactly what a reader would inherit without noticing, and each met a different fate in the second session. The first response added a checkpoint header dated 1 October 2026. The prompt dated the two source notes to that day but never supplied a checkpoint date, and the request occurred on 2 October, so the header was an unsupported assumption in the first output. It entered the second session only as part of the pasted input; the second response did not repeat it, using its own header dated 2 October instead.

The first response also strengthened supplied wording: a confirmed budget and launch date came back as fixed and locked decisions, descriptions stronger than what was given. Those labels passed into the second request inside the pasted checkpoint; the record attributes them to the first response.

Finally, the second response labelled the two new entries S3 and S4 as supplied on 2 October 2026. That matches the request's local calendar date, but the prompt did not explicitly date those notes, so they are inferred receipt dates, not verified publication metadata. With no third request, the record contains no further instance of them.

What does this not show?

On the memory side, the separation holds throughout. SentX's homepage describes dynamic symbolic memory of people, preferences, and shared conversations, with important memories strengthening and lower-value details fading. Its privacy policy states that submissions are not confidential, that a shared memory system can surface content across users, and that account-level deletion does not reach information already incorporated into training, model weights, or shared memory. None of that was exercised by this demonstration, which deliberately resupplied everything. The research summarizer page likewise documents structured summaries while warning that summaries can contain mistakes and key claims need checking against originals. These are first-party descriptions, not measured guarantees of recall.

Repetition, or consistency with the supplied notes, does not independently verify what those notes contain. The habit for anyone resuming work this way is to check dates, decision strength, and sources alongside the arithmetic before building on a checkpoint. The scope limits hold throughout: one synthetic project, two responses, no benchmark, no accuracy claim, no performance comparison, and no executed actions.