NeuroVault

NeuroVault is a local-first AI memory system that fixes the central failure of LLM agents: they forget you after every conversation. Built as a Tauri 2 desktop app, NeuroVault stores notes as plain markdown in your home directory, indexes them with a Rust-powered hybrid retrieval engine (semantic + BM25 + knowledge graph), and exposes them to Claude over MCP. Most agent-memory products today are RAG pipelines in disguise. NeuroVault treats memory as a structured, updatable, inspectable knowledge base - what you get when you stop chunking-and-praying and start treating memory as a living wiki.

Architecture

+-------------------------------------------------+
|  Tauri 2 desktop app (React 19 + TypeScript)    |
|  Editor / Graph / Compile / Sidebar / Palette   |
+-----------------------+-------------------------+
                        | Tauri commands  +  HTTP :8765
+-----------------------v-------------------------+
|  In-process Rust backend                        |
|  - axum HTTP server (the MCP proxy talks here)  |
|  - hybrid retriever (semantic + BM25 + graph)   |
|  - fastembed-rs (BGE-small ONNX, local)         |
|  - notify file watcher                          |
+-----------------------+-------------------------+
                        | SQL + vec0
+-----------------------v-------------------------+
|  SQLite + sqlite-vec  (~/.neurovault/...)       |
|  brain.db, vault/*.md, raw/, assets/, cache/    |
+-------------------------------------------------+

Key Features

Why this is not RAG

RAG is an answer-pipeline: question → chunk → embed → retrieve → stuff context → generate → repeat. The corpus is dead data. Contradictions are invisible. Provenance is a prayer. NeuroVault is a knowledge layer - it differs from RAG in five specific ways that map directly to what a living internal wiki needs.

• Accumulates via Ebbinghaus strength decay + access reinforcement (RAG re-chunks)
• Structured with Karpathy's 3-layer raw/wiki/schema pattern (RAG is flat chunks)
• Three automatic link types: semantic similarity, shared entities, explicit wikilinks (RAG has none)
• Silent fact capture promotes casually-dropped facts with wiki-link provenance (RAG cites the chunk)
• Temporal fact tracking - when new facts contradict old, supersession + recency penalty stops pollution (RAG cannot challenge or update)

Performance benchmarks

All numbers are reproducible locally. Run cd server && uv run python ../benchmarks/run_recall.py to verify. The benchmark uses 25 hand-crafted notes and 25 queries (5 easy, 10 medium, 10 hard).

• Hybrid retrieval: 92% Top-1, 96% Top-3 hit, 0.94 MRR, 73ms median latency
• With cross-encoder rerank: 92% Top-1, 100% Top-3, 0.96 MRR, 133ms latency
• Hard queries (no keyword overlap): 9/10 top-1 without reranker
• Embed a note: ~20ms · Full vault ingest (25 notes): ~4s cold start
• Tokens per answer: ~275 flat regardless of vault size (paste-whole-vault baseline: 93k+ growing linearly)

The silent fact-capture pipeline

The killer feature: drop a fact in conversation and NeuroVault picks it up without you saying "remember this." A UserPromptSubmit lifecycle hook pipes every prompt through a regex extractor that recognizes 8 patterns: preferences, decisions, stack choices, deadlines, identity, anti-preferences, deployment targets, and explicit "remember that..." callouts. End-to-end bench: 80% Hit@1, 100% Hit@3 on 15 paraphrased questions that never use original wording.

Tech stack rationale

Each layer was chosen for local-first guarantees: Tauri 2 over Electron (~30MB vs ~200MB installed), Rust backend in-process (no Python sidecar at boot), sqlite-vec for KNN in pure SQL (no separate vector DB), fastembed-rs for ONNX-quantized embeddings (no API keys, no internet), MCP as the agent contract (works with Claude Desktop, Claude Code, Cursor, any MCP-speaking client).

API Endpoints