Long campaign AI RPG.

The mile-marker problem

Take a campaign at turn 500. Your character has built a small business, made allies in two factions, killed a rival merchant in the back of a ship, raised two apprentices, and lost one of them in a winter that should have been mild. None of that fits in a context window. None of it can be summarised into a sentence without losing the texture that made it matter.

The standard AI-context approach erodes everything but the most recent turns. Your apprentice's name vanishes. The rival merchant comes back to life. The winter never happened. The story is a sandcastle in a slow tide.

Three architectural moves that change the math

1. Mechanical state in a real database. Coins, inventory, NPCs, factions, locations, properties — all in Postgres rows, not chat history. The AI narrates around the database; the database is the truth. Inventory math survives the long haul because it isn't narrated, it's queried.

2. Hierarchical chapter compression (RAPTOR).Every chapter compresses into a tight summary. Summaries compress into act-level summaries. Act-level summaries compress into campaign-level summaries. The hundredth turn can pull a relevant detail from chapter two without flooding the context window.

3. Semantic vector retrieval of episodic memory.Beyond chapter summaries, individual important moments are embedded. When the AI is composing a turn at session 50 and the current scene mentions an old promise, the system can pull the specific exchange where that promise was made even if it was 800 turns ago.

What this gets you

• Characters who keep their names, voices, and relationships across hundreds of turns.
• NPCs who remember specific exchanges from chapters earlier — "you said you'd come back by spring; you didn't."
• An economy that doesn't reset. Prices shift, supply chains break, factions actually decline.
• A character sheet that grows, not drifts.
• Save / restore that actually restores the world, not a vague summary of it.

Honest limit

Long-campaign coherence isn't magic. The narrator is still an LLM and can hallucinate small details (especially flavour text, ambient NPCs, off-screen rumours). The mechanical truth is locked down; the prose is not. If you're running a 1000-turn campaign and the AI gets a minor detail wrong, you can OOC-correct it and the engine will adjust. We can't promise zero drift. We can promise the database is right.