In this guide
A long context window is not a campaign ledger
OpenAI lists a 1,050,000-token context window for GPT-6.2 Astra in the API. That gives you room to supply relevant material, but it does not settle which version of an old event is true or which character witnessed it. A compact ledger makes those decisions explicit before you hand a scene to an AI game master.
The canal guild publicly promised to reopen the north lock at dawn. Only its archivist saw the copper seal break the night before. Dockworkers say a rival broke it; nobody has confirmed that. If the GM tells the whole council that the rival is guilty, the campaign has already changed, regardless of how much session history the model read.
Store facts, promises, and rumors separately
Write each entry with a status and a source. ‘Public promise: north lock opens at dawn; council minutes, session 8.’ ‘Private observation: copper seal broken; archivist only, session 9.’ ‘Rumor: rival blamed; dockworkers, unverified.’ Add the current time and outstanding deadline at the top, and keep the earlier entry when a later scene adds a correction.
When briefing the next scene, supply those lines and ask the model to say what the speaker could know, then give the in-character reply. If a guild envoy has never seen the seal, they can repeat the rumor or call for an inspection; they should not describe the break as firsthand evidence. Keep the check out of the player-facing prose so it does not spoil the private note.
Test the reveal before you trust the recap
Replay the council scene twice: first before the archivist testifies, then with one new entry stating that she shows the broken seal in public. After the reveal, the council can know it broke. It still cannot know who broke it. If those two states produce the same response, shorten the brief or make knowledge boundaries more explicit before changing models.
When a scene ends, review proposed changes before adding them to canon. This is an illustrative workflow for evaluating Astra, not a claim about automatic campaign memory or model availability in Playworlds. You can use the same ledger at a human-run table.

