Nerve

Companies run on knowledge trapped in people's heads and scattered across docs, tickets, chats, and calls. When someone leaves or a process changes, that memory walks out the door, and no one can trust that the documentation matches how work is really done.

Nerve builds a company's operational memory automatically. A lightweight agent observes how real work happens and ingests existing sources (docs, tickets, chat, call recordings, code), then distills it into two living assets: human-approved step-by-step walkthroughs, and a knowledge base where every stored fact links back to its evidence. People ask questions in plain English and get cited answers; other AI tools query the same memory over MCP while they work.

Claude Opus handles reasoning and synthesis, Haiku the high-volume passes, and Claude vision turns work screenshots into factual narratives. Evidence is retrieved with hybrid search (vector + full-text + Reciprocal Rank Fusion over Postgres/pgvector, plus knowledge-graph expansion) using local BGE embeddings, then synthesized into streaming answers whose citations are mechanically validated against their sources, so nothing is stored or claimed without provenance. A read-only MCP server exposes the memory to external agents.