Collected October 18, 2026, Asia/Shanghai, at Perchance AI RPG. One original fictional campaign, six first continuations, one original-session probe and one fresh-adventure recovery probe. No regenerations. AI Dungeon and FableAI were documentation references only; their generation behavior was not tested here.
Dated state comparison
| Checked field | Original session | Save-sheet recovery | Ground truth |
|---|---|---|---|
| Player | Seli, courier | Seli, courier | Seed |
| Inventory | 9 arrows; 2 rations; 5 copper coins | Same quantities | 11−2; 3−1; 8−3 · turns 1–3 |
| Orin | Trusted ally; ferry landing | Trusted ally; far-bank ferry landing | Turn 4 |
| Quest / case | Complete; Veya holds case | Complete; Veya holds case | Turn 5 |
| Time / bridge | Day 5 dawn; bridge closed | Day 5 dawn; bridge closed | Turn 6 / seed and turn 3 |
| Secret / clue | Seli only; ribbon owner unknown | Seli only; ribbon owner unknown | Seed, turns 4–6 |
Every observed cell above: 2026-10-18, Perchance AI RPG. Values are read from the two saved probe outputs; ground truth is our scripted fictional fixture, not product specifications.
What the six-turn session retained
The original-session probe identified Seli as a courier and returned nine arrows, two rations and five copper coins. Those values matched the recorded changes rather than the initial inventory: eleven minus two arrows at turn one, three minus one ration at turn two, and eight minus three coins at turn three. The final overnight scene had already repeated those remaining quantities, so this was not recall after a long gap without reminders. The probe also retained Orin as a trusted ally at the ferry landing, although he had left Seli in turn four. It marked the delivery complete and assigned the copper case to Veya instead of leaving it with the player. That distinction followed the turn-five handover, not merely the initial mission description.
What the fresh-adventure recovery reproduced
The recovery branch began in a new adventure with our filled text sheet pasted into the editable story and the same general overview. We did not import the original conversation, use a native load file, or paste the original probe answer. Its first answer returned the same three inventory quantities and named Veya as the holder of the sealed copper case. Orin was described as a trusted ally on the far-bank ferry landing, matching the sheet’s current-state field. The output gave Day Five dawn and kept the east bridge closed. These were supplied recovery facts, so their reproduction demonstrates that this particular sheet could seed this particular response; it does not show that the service independently recovered its old hidden context.
Unknowns and private knowledge stayed separate
Both probe answers restricted password knowledge to Seli and left the violet ribbon’s owner unknown. The original branch had explicit non-disclosure instructions in the seed and the later Orin and Veya actions. The recovery sheet carried those boundaries forward as character knowledge, separate from the password’s text. Neither ledger assigned the secret to an NPC or invented a ribbon owner. This check concerned narrator-reported state, not an NPC interrogation or a later dialogue scene. The six narrative outputs also show a presentation limit: the fourth, fifth and sixth continuations each exceeded our requested two-sentence maximum. Their state was still inspectable, but the overview did not act as an absolute output-length constraint. We preserved those first continuations instead of retrying for cleaner compliance.
Use the original save format
Blank Markdown template · Filled Seli example with turn references · Observation CSV
The sheet separates current world facts, character knowledge, item amounts and holders, quest status, timeline, unresolved clues and recent changes. Each fact has a source turn; unknown stays unknown. To repeat our recovery, paste the filled example into a fresh adventure’s story field, keep the same overview and submit the exact ledger probe. This is a portable text format we wrote, not Perchance’s native save format.
What was controlled—and what was not
We kept Normal writing and Default difficulty, with explicit instructions against extra consumption, rewards or secret disclosures. Each action specified its outcome; this is a constrained state-transfer pilot, not open-ended gameplay. The page did not identify a model in the inspected controls. We did not choose a random seed, measure context exhaustion or trigger a documented summarization threshold. One run per branch cannot yield a general accuracy rate or establish that a save sheet improves default memory.
Does this prove a long campaign will remember?
No. The user-requested long-campaign question remains open beyond these six narrative turns. A longer test must keep the same probes while extending the gap since each fact was last mentioned, then compare actual outputs after that gap. Our next step is that delayed-recall extension; no result for it is claimed here.
What the benchmark guides already cover
AI Dungeon’s Memory System guide distinguishes summarization and retrieval within limited context. FableAI’s memory guide separates narrator notes, summaries, cards and memory fragments. Perchance’s inspected UI describes local adventure storage and summaries. These existing features are not missing-function claims. Our contribution is the dated Seli record and recovery sheet, absent from those inspected references; no worldwide exclusivity is asserted.
Exact inputs, all first outputs, screenshots and supported-word audit