How Our AI Dungeon Master Now Remembers What Was Said 45 Rounds Ago
An AI Dungeon Master rarely forgets that something happened. It forgets how it was said: the price the ferryman named, the exact words of an oath, the name a stranger gave at the door. Since October 2026, LoreKeeper keeps every round word for word and hands the DM the passages that matter, turn by turn. In our benchmark, facts from 45 or more rounds earlier now reach the DM 63–79% of the time, up from 12–19%.
Why do summaries lose the details players care about?
Because a summary keeps events and drops wording. An AI model can only read so much text per turn, so older rounds get compressed: we explain the mechanism in why AI Game Masters forget your campaign. “The party hired a ferryman” survives the compression. “Three silver, and never after dark” usually does not.
LoreKeeper already had structured memory for the big picture: the Codex living wiki keeps one article per NPC, place and thread. That tells the DM who the ferryman is. It was never meant to hold every sentence he said, and players notice exactly those sentences when the DM gets them wrong.
How does verbatim round recall work?
Every finished round is stored word for word, and each turn the DM receives the few passages from earlier rounds that bear on the present moment. Step by step:
- Every round is kept. What each player said or attempted and what the DM narrated are saved as a transcript and split into short passages.
- Your action drives the search. When you act, the passages are searched by meaning and by exact words using what you just did or asked. The scene adds its own leads: the NPCs present, the place you are in and your active quests.
- A judge keeps only what matters. Each candidate is checked against one question: does it bear on what is happening now, or state a name, promise, number or event the narrator must stay consistent with? Only those pass.
- The DM gets a short, labelled record. At most four passages, about 4,000 characters, marked as the verbatim record. What players said is labelled apart from what the DM narrated, so a player's claim is never mistaken for established fact.
Recent rounds are not searched: the DM already reads them in full. Recall covers the stretch that used to exist only as a summary.
How much better does the DM remember?
We measured it before switching it on. The benchmark takes concrete facts said in rounds 3 to 15 of real campaigns that ran for 50 rounds or more, then checks whether each fact reaches the DM when it comes up again 45 or more rounds later. Each fact is tested three ways: asked directly, only implied, or simply needed for the DM to stay consistent.
| Fact from 45+ rounds earlier | Reached the DM |
|---|---|
| Previous memory layers | 12–19% |
| With verbatim round recall | 63–79% |
The high end is facts the player asks about directly (79%); the low end is facts the DM has to keep consistent without being asked (63%). To be precise about what this measures: it is whether the fact is in front of the DM when it narrates, which is the precondition for getting it right. It is not a promise that every answer is perfect.
What does that look like at the table?
An illustrative example of the kind of moment recall is built for:
- Round 8. The ferryman agrees to take the party across for three silver, “and never after dark”.
- Round 60. The party comes back to the river at night and asks for a crossing.
With a summary, the DM knows there is a ferryman and may happily row you across. With verbatim recall, your request and the ferryman present in the scene both point back to round 8, and the DM has his own words in front of it: he refuses, or names a much higher price for breaking his rule.
What changes in the Codex?
The Codex is now written from the same round transcripts. Before, it compiled from shorter notes that often did not exist for the rounds it was updating, so articles could miss what had actually been said. Now each article is built from what was said and done in those rounds. If you want the bigger picture of how the wiki and the DM's memory fit together, see our LLM wiki for RPG campaigns.
Which plans get the improved memory?
All of them, free included. Verbatim round recall is on for every campaign and there is nothing to switch on. Long campaigns benefit the most, because they are the ones with the most rounds outside what the DM reads in full. The Codex living wiki stays part of the Hero and Legend plans.
Frequently Asked Questions
Does an AI Dungeon Master remember what happened earlier in the campaign?
It depends on the platform. Most AI Dungeon Masters only see the most recent messages plus a summary of older ones, so exact details fade. Since October 2026, LoreKeeper keeps every round word for word and, on each turn, brings back the passages from earlier rounds that bear on what is happening now.
How much better does LoreKeeper remember with verbatim recall?
In our internal benchmark, built from facts said in rounds 3 to 15 of real long campaigns and asked about 45 or more rounds later, the fact reached the AI Dungeon Master 12–19% of the time with the previous memory layers. With verbatim round recall it reaches it 63–79% of the time, depending on whether the player asks directly, implies it, or the DM simply has to stay consistent with it.
Is verbatim round recall the same as the Codex living wiki?
No. The Codex is a curated wiki with one article per NPC, place and thread, which is great for who someone is but summarises away exact wording. Verbatim recall brings back the actual passages: the price the merchant named, the words of the oath, what your character promised. They work together, and the Codex is now written from the same round transcripts.
Which LoreKeeper plans include the improved DM memory?
All of them, the free plan included. Verbatim round recall is on for every campaign and there is nothing to switch on. The Codex living wiki, which now compiles from the same transcripts, remains part of the Hero and Legend plans.
Does the AI Dungeon Master read the whole campaign every turn?
No. Reading everything would be slow and would bury the important detail in noise. Each turn the DM receives at most four short passages from earlier rounds, about 4,000 characters, chosen because they bear on the current moment or on a fact the narrator must keep consistent.
Start a Campaign That Keeps Its Word
Every round kept word for word, on every plan. Free to start, up to 6 players, no credit card.
Play Free →Related guides
What's New in LoreKeeper (October 2026)
Scene images built from the round, NPC portraits, talking NPCs, smarter companions, a narrator selector and clearer combat.
Karpathy's LLM Wiki, But for Your D&D Campaign
RAG, LLM wiki, or autonomous skills? How LoreKeeper runs the retrieve + compile layers in production for an AI Dungeon Master.
Why AI Game Masters Forget Your Campaign
The context-window mechanism behind AI memory drift, and the retrieved store — dossiers, checkpoints, source resolver — that fixes it.