How Memory Should Work in a Persistent AI RPG

Transcript is evidence, not state

A transcript records words in sequence. It does not reliably answer whether the gate is open, whether a promise was fulfilled or which companion privately learned the king's secret. Asking a model to reread an ever-growing transcript also becomes slow and inconsistent.

Structured state turns important outcomes into explicit records: quest phase, location, inventory ownership, injury, faction standing, relationship milestone and known fact.

Different memories belong to different minds

The player, Game Master and each companion should not have identical knowledge. A secret heard by one character remains private until they share it. A rumor can be stored as uncertain while an observed event becomes established fact.

This separation prevents accidental spoilers and gives investigation real meaning. It also lets companions form conflicting interpretations without the system losing track of objective world state.

Checkpoints protect canon

Speech recognition and generative systems can make mistakes. Before a consequential action is sealed, the player should be able to correct a misheard sentence or report a canon violation. After the checkpoint, legitimate outcomes should remain durable.

That boundary distinguishes error recovery from rewinding a failed choice. A persistent campaign needs both fairness and consequence.

Memory should change the next scene

Stored facts matter only when they influence play. A saved village alters later rumors, a wounded companion changes the route and a political betrayal affects access to a port. The system should retrieve a small set of relevant memories for each scene rather than dumping the entire campaign into every response.

Living Party stores durable world memories alongside relationships, quests and route state so the Game Master and companions can respond to the campaign the player actually created.

Frequently asked questions