Overview
What this is
The compact read before the technical details.
A brain-inspired semantic memory system that stores, deduplicates, and retrieves memories by meaning rather than keywords. SQLite is the source of truth with LanceDB for vector search and Transformers.js for local embeddings. Every memory activates across 11 brain regions in a 3D atlas built with Three.js. Ships as both an npm CLI and a stdio MCP server for AI agent integration.
Capabilities
What it actually does
The useful parts, pulled out of the paragraph wall.
Four-stage write pipeline — LLM extracts entities, relationships, region weights, and emotional tone; local sentence-transformer embeds; semantic deduplication scores candidates; SQLite + LanceDB persist atomically
Published npm package (atlas-mcp) exposing 9 MCP tools (add, search, get, related, entities, update, delete) for persistent agent memory via any MCP-compatible client
3D brain visualization maps every memory to 11 cortical and subcortical regions with bilateral hippocampal hemispheres, rendered in Three.js with force-graph exploration
Conservative entity linking avoids false merges — ambiguous matches stay separate, related-memory links are derived at read time from shared entities plus vector similarity
Implementation
Technology with jobs attached
Names are less useful than responsibilities. This is what each piece is doing.
Node.js / Express
HTTP API and MCP stdio server with ESM throughout
SQLite
Source-of-truth memory store with typed nodes, entity links, and region activations
LanceDB
Local vector index for semantic search over 384-dim sentence-transformer embeddings
Transformers.js
On-device embedding model — no cloud dependency for vector operations
React + Vite
Frontend SPA with TypeScript, routes for Atlas, Catalog, Compare, and Graph views
Three.js
3D brain atlas visualization (~3600 lines) with region activations and memory footprints
MCP
Stdio server protocol exposing memory tools to Claude Desktop, Cursor, and other agents