Synapse - Brain RAG & Agent Memory

Synapse is the knowledge layer: it turns a corpus of decisions, project docs, and business notes spread across nineteen repositories into something an agent can actually query. Documents are chunked, embedded locally through Ollama's mxbai-embed-large — no corpus text leaves the machine — and stored in Postgres with pgvector. Retrieval runs over that index and answers through a document-QA workflow on an event-driven FastAPI and Celery pipeline. The part worth arguing about is the measurement. Retrieval systems degrade quietly: a change improves one query class and silently breaks another, and without a harness you find out months later. Synapse carries a golden set, a tracked eval run history, and a corpus fingerprint guard that refuses to compare two runs taken against different corpora — an incomparable verdict instead of a misleading improvement. Query-log mining reads the real query log and classifies gaps into abstained, low-confidence-answered, and suspected-confidently-wrong, which is how new golden cases get proposed from live traffic rather than from imagination. The memory layer models peers, episodes, and facts, with extraction at ingest time, consolidation on a schedule, confidence decay, and explicit contradiction resolution — so a fact that was true last quarter is superseded rather than silently overwritten. One deliberate boundary: Synapse owns the semantic half of the Brain. The structural half — graph queries over the wikilink structure between documents — lives in the Rust Console instead. Semantic similarity and structural adjacency answer different questions, and collapsing them into one index makes both worse.