Can you trust an AI to remember your D&D campaign?

Here is the moment that makes this matter. You make a ruling, and a player is certain it went another way. They swear the baron promised them the land, you remember it as a maybe, and the table stalls while you relitigate a conversation from two months ago. Nobody enjoys that, and however you call it, someone can walk away feeling the game cheated them.

With a recorded campaign you can settle it in seconds, in front of everyone. Ask the question in Discord at the table, and Recap Raven shows the actual line from the session, quoted from the transcript. The disagreement ends because the record is right there, not because you pulled rank. A player seeing the real quote and nodding is the kind of trust I could not find in any other tool, and it is a big part of why I built this one.

The common approach is simple: take the whole session transcript, hand it to a chatbot, and ask for a summary or an answer. It holds up until it does not. A long campaign is far more text than an AI can read at once, so details fall out, and when it cannot find something it tends to fill the gap with a confident guess rather than admit the gap.

The result is canon you cannot trust. Two sessions blur together, a name comes back slightly wrong, an event that never happened gets stated as fact. The deeper problem is that you cannot check any of it, because the answer points to nothing. You either take it on faith or dig through the transcript yourself, which is the work you were trying to avoid.

The most important property of a memory you can trust is that every answer shows its source. Recap Raven quotes the actual transcript behind each answer, so you can open it and read the line for yourself. When a ruling hangs on a detail, that thirty-second check is the difference between guessing and knowing.

Citations also keep you honest about what the tool is. Treat it as a cited assistant that is usually right, and verify when it counts. The quote is there precisely so you can confirm it and move on, instead of wondering whether the answer was real.

A memory you can trust has to be willing to come up empty. Recap Raven abstains when the transcript does not support an answer, rather than inventing one. That sounds minor, and it is the thing that protects your canon. A tool that always answers will answer even when it is wrong, and you carry that error into the next session.

Saying nothing when there is nothing to say is a feature worth looking for. It tells you to go check, or to make the call yourself, instead of quietly handing you a fact that was never true.

On top of the citation, every answer comes with a confidence band: high, medium, or low. It is derived from how well the transcript actually backs the answer, not the AI rating its own certainty, which is the one number you cannot trust it to give you honestly. A high band means the evidence is solid. A low band is your cue to open the quote and check, or to make the call yourself.

Taken together, the citation, the abstention, and the confidence band tell you how much weight to put on any answer at a glance. You are never left guessing whether the tool is sure of itself or bluffing.

Under the hood, Recap Raven does more than forward your transcript to a chatbot. It builds a memory: it finds the relevant moments across your whole campaign, ranks them, and answers from those, with the quotes attached. That retrieval work is where the accuracy comes from, and it is why a question about session three still finds the right answer thirty sessions later.

I spent real engineering effort there on purpose, because a summary you cannot trust is not worth having at the table. Player-safe recaps and GM-only notes stay separate as well, so checking your canon never leaks a secret to the players.

None of this makes the answers perfect. AI gets things wrong, and Recap Raven will too. The point is to build for that reality: cite everything, band each answer by how well the evidence backs it, admit uncertainty, and keep the source one click away, so a wrong answer is easy to catch instead of quietly believed.

That is the standard I wanted as a DM, and the one I hold the tool to. I needed a campaign memory I could run a real game from, where I can trust an answer because I can see exactly where it came from.