From Meeting Protocols to Corporate Memory

Meetings produce decisions, owners, and dependencies that mostly vanish into protocols nobody rereads. Nobody can easily answer "why did we decide this?" or "what changed since March?" — full-text search and vector RAG find similar text, not relationships or history. Build a pipeline that turns meeting protocols into a queryable memory. The approach is open: knowledge graphs, temporal graphs, GraphRAG, hybrid setups, or your own idea. Focus on three things: what can be extracted automatically (entities, decisions, owners, timelines), how humans review and correct it without too much friction, and how the data model handles renamed entities, superseded decisions, and time-based questions.

We're an energy supplier (EVU) operating under a regulatory environment that changes constantly: the revised StromVG, new ElCom rules, recurring tariff and reserve obligations. Each change forces decisions that later need to be defended, revisited, or reversed. Right now, tracking what was decided under which rule lives only in people's heads. Decisions get argued over again, contradictions surface too late, and knowledge leaves with staff turnover. A working corporate memory would let us trace why a decision was made, see what depends on it, and answer to management or regulators with confidence.

Generating a dataset is part of the challenge. Create synthetic meeting histories for a realistic organization over time, with edge cases built in: decisions that get quietly revised, projects renamed mid-stream, unclear ownership, action items that resurface. Make the data hard on purpose: it should stress-test your pipeline, not just demo it.