Summary

Where it earns its place is the folder, not the snippet.

Plenty of servers give an assistant a place to remember things. This one is built around dropping a directory into the index — images, CAD, PDFs, data files and source all get their own metadata extraction — and quick_load makes that cheap enough to do for a folder you only care about for an hour.

What it is

An MCP server that keeps an assistant's memory in a ChromaDB instance running on your own hardware. Beyond storing snippets, it walks directories, extracts metadata per file type, and makes the result searchable — nothing leaves the machine.

What you get

Memory tools — `store_context`, `search_context`, `list_collections`, `find_similar_patterns`
Bulk ingest — `scan_directory` to preview what is there, then `batch_ingest`, `ingest_file`, or `quick_load` for a throwaway collection you drop again with `unload_collection`
Collection management — `get_collection_info`, `export_collection` and `import_collection` for JSON backups, `list_file_types`
Photo metadata — `extract_exif` pulls camera, lens, exposure, GPS and date taken out of JPEG and TIFF files
Watched folders — `watch_folder`, `stop_watch`, `list_watchers` auto-ingest anything dropped in
Duplicate detection — `find_duplicates` with partial, full or perceptual hashing, plus `compare_files` and `find_collection_duplicates`
Six collections created by the setup script, including code_snippets, configurations, troubleshooting and project_docs

Requirements

Bun as the runtime, Docker to run ChromaDB, and an MCP client. The documented setup runs the chromadb/chroma image with host port 8001 mapped and a persistent volume, then `bun install` and `bun run setup` to create the collections. Your client config points at index.js with CHROMADB_URL set to the ChromaDB address. MIT licensed.