Campaign memory
An answer you can check
- 65,000+
- Development AI requests
- 2,500+
- AI quality evaluations
- 200+
- Real campaign questions in a quality benchmark
Development requests cover the 12 months to October 2026. Quality evaluations count individual answers assessed across different model families, versions, reasoning budgets and task approaches, including repeated questions with different settings. Question sets can vary between benchmarks.
We regularly evaluate how well Recap Raven preserves characters, decisions and events, and how it handles questions the available record cannot answer. At Recap Raven, we believe AI should abstain when the evidence is insufficient, rather than give an unsupported answer. When we change something, we check the results against the current benchmarks.
Privacy comes first when we choose AI providers. We only use providers that do not train on your campaign content. We actively test new AI advances against our campaign benchmarks, checking whether they improve the work Recap Raven does for your table before adopting them.
Searchable Memory gives you references to the material behind an answer. Session citations let you check the surrounding conversation, giving you more confidence that the answer reflects what was actually said. Our architecture and testing give particular attention to distinguishing plans discussed at the table from events that actually happened. An abandoned idea should never become part of your campaign’s history.
Click to inspect the source quote.
Sources help you check an answer, and editing tools let you correct the campaign record. Read more about checking campaign memory.
What we look for when evaluating AI
We check names, relationships, important decisions and continuity across sessions, including questions that require more than finding a single detail.
Evidence you can use at the table
A citation gives you a place to investigate. Open it, read the surrounding exchange and decide whether it supports the answer. Transcripts can contain errors too, especially around unfamiliar names or overlapping voices. Your knowledge of the game remains important.
You can correct the campaign record when something is wrong. Recap Raven supports automatic recap publication, or a review workflow where you check and edit recaps before sharing them with your table. Choose what suits you, whether you want less admin between games or prefer to review every detail first.
Quality is ongoing work
Our engineering checks cover security, permissions, campaign isolation, accessibility and product workflows. AI evaluation asks a different question: whether the answers accurately describe your game. Campaign audio, names and source material vary, so we do not publish a universal accuracy percentage.