How does an AI Dungeon Master remember what happened last session

The short answer

An AI Dungeon Master doesn't remember your session the way you remember a conversation. It reconstructs a working picture of it every turn, drawing from several layers of storage that each hold a different kind of memory. Fabulator splits this into three broad categories: Canon (the stable facts of the world), State (what's true right now), and Memory (what happened, decaying over time the way a human memory would). Understanding those layers explains both why a good AI game master feels consistent and why it occasionally forgets something you'd expect it to know.

Canon: the facts that don't change mid-session

Canon is the immutable backdrop. Faction names, magic system rules, the history of a kingdom, a character's stated backstory. Once established, this layer stays fixed unless a player deliberately edits it or the story canonizes a new fact through play. When you ask "wait, didn't the blacksmith mention a rebellion twenty years ago," the narrator is pulling that answer from Canon, not guessing at it.

This is also the layer most resistant to drift. A narrator that contradicts its own world's history breaks immersion fast, so Canon is treated as close to ground truth as the system allows.

State: what's true this turn

State is the live scoreboard: your inventory, your relationship standing with an NPC, whose turn it is, active graph edges connecting characters and locations. It updates through validated deltas, meaning a proposed change (you picked up a sword, an NPC's disposition shifted) has to pass a check before it's committed. This is why an AI game master can say "you can't equip that, your strength is too low" instead of just narrating around a stat it never actually tracked.

State answers the question "what does the world look like right now," and it's rebuilt fresh at the start of every turn rather than inferred from the conversation history alone.

Memory: the part that actually decays

This is the layer most people mean when they ask how an AI Dungeon Master "remembers." It's an append-only record of what happened, plus a rolling summary and vector clusters of past events, and it genuinely decays over time. A conversation from ten turns ago is easy to recall in detail. A conversation from two hundred turns ago has usually been compressed into a summary, or reduced to a handful of significant events, the way a person remembers the shape of a trip rather than every meal they ate on it.

That decay isn't a bug. Every turn has a limited token budget for context, and stuffing the full history into every prompt would be both expensive and, past a certain length, actively worse for narrative quality. The system has to choose what's worth keeping in detail and what can be summarized, and it errs toward keeping anything flagged as significant (a character death, a major plot turn, a promise made to an NPC) even as smaller details fade.

Why this sometimes feels wrong

If an AI game master forgets a minor detail from three sessions ago, it's usually because that detail was never significant enough to survive the decay curve, not because the whole system is unreliable. The failure mode people actually notice is different: it's when something that should have been Canon or State (a stated fact, a tracked item) gets treated as if it only lived in Memory, and Memory's job is recall, not guarantee.

A well-built AI Dungeon Master is judged less by how much it remembers and more by whether it remembers the right things: the facts that must never change, the state that must always be accurate, and the handful of moments from your story that actually mattered. Most complaints about "the AI forgot" trace back to one of those three layers doing the wrong job, not to memory failing in some generic sense. Once you know which layer is responsible for which kind of fact, it's much easier to tell whether a given slip is a real bug or just the expected cost of a system that has to summarize, not transcribe, everything that happened.

Try it yourself

The easiest way to see this in practice is to watch it work across a real session rather than take a description of it on faith. Fabulator's World Seeds let you drop into a published setting and play a few turns for free, no install required, and you'll notice the difference between "the world remembers the rules" and "the world remembers your choices" pretty quickly once you're in one.