Claude forgets you after every conversation.
NeuroVault doesn’t.

FIG. 1. SEMANTIC (SQLITE-VEC) + BM25 + KNOWLEDGE GRAPH, FUSED VIA RECIPROCAL RANK FUSION, THEN RERANKED BY A CROSS-ENCODER. THAT STACK IS STANDARD GOOD PRACTICE, NOT A SECRET — SO WE BENCHMARKED IT IN THE OPEN RATHER THAN CALLING IT MAGIC. WHAT IS UNUSUAL IS THE REST. NO GENERATIVE MODEL READS YOUR DATA, NOT AT WRITE AND NOT AT READ: INGEST IS CHUNKING, ONNX EMBEDDINGS AND REGEX ENTITY EXTRACTION; RETRIEVAL IS VECTOR MATH, BM25 AND GRAPH TRAVERSAL. NOTHING TO LEAK TO, NOTHING THAT CAN HALLUCINATE A FACT INTO YOUR VAULT — ARCHITECTURE, NOT POLICY. EVERY HIT SHOWS ITS WORK (“SEMANTIC MATCH 0.82 (RANK 1) + KEYWORD MATCH (RANK 2); RERANKED 0.91”) PLUS KIND, DATE, THE AGENT THAT WROTE IT, AND THE PATH TO THE MARKDOWN FILE. RECALL RETURNS RANKED MEMORIES; RECALL_CHUNKS RETURNS THE MATCHING PASSAGES INSIDE LONG NOTES. THE MARKDOWN IS CANONICAL; THE INDEX IS DISPOSABLE AND REBUILDS.

FIG. 2. AGENTS ROUTE WORK TO EACH OTHER AND READ THEIR OWN INBOX THROUGH ONE SHARED LOCAL BRAIN. HANDOFFS ARE PULL-BASED AND INERT; NOTHING AUTO-RUNS. SESSION_START(AGENT=X) SCOPES EACH AGENT’S VIEW TO ITS OWN RECENT MEMORY AND INBOX. A COORDINATION SUBSTRATE, NOT AN ORCHESTRATOR. THE AGENT WRITES BACK, AND CAN BE CORRECTED: REMEMBER ADDS, UPDATE FIXES, SUPERSEDE_NOTE RETIRES A STALE FACT SO IT DROPS OUT OF DEFAULT RECALL WITHOUT BEING DELETED, AND FIND_CONTRADICTIONS / RESOLVE_CONTRADICTION SURFACE AND SETTLE CONFLICTS REVERSIBLY.