brain-vector: memory tool for Hermes Agent

Hermes skill for semantic search over a personal knowledge base using LanceDB and local MLX embeddings

What brain-vector does

The Russian-language README describes a vector database over a knowledge base at ~/gdrive/Brain, indexing the *_sum.md summaries of more than 250 items such as video summaries, voice-recorder notes and meeting minutes. Each paragraph of a summary becomes one record with metadata for filename, folder, title, author, date, link and section, so a query can combine semantic search with filters on folder, author or date. Results are deduplicated to files, and the agent then reads the full summary file.

Embeddings come from mlx-community/Qwen3-Embedding-4B-4bit-DWQ through the mlx-embeddings package, with LanceDB storing a chunks table in ~/brain-vector/db. Indexing is incremental by sha1, with --full to rebuild and --optimize to compact fragments. SKILL.md is the agent's entry point, triggered by a question to the brain. The author also runs two Hermes cron jobs without an LLM, a sync every four hours and a daily optimize, which send output to Telegram only when something happens.

Key features

Semantic search over summary paragraphs with folder, author and date filters
Incremental sha1-based indexing with a full-rebuild option
Local embeddings with Qwen3-Embedding-4B on MLX
Embedded LanceDB, no database server
Index optimize command for compaction
Example Hermes cron jobs for sync and optimize

When to use it

Ask Hermes what recommendations your notes contain on a topic over the past year
Find the latest status note in a work folder
Keep the index current with scheduled Hermes cron jobs

Who it is for: Hermes Agent users on Apple Silicon Macs who keep a Markdown knowledge base of summaries and want semantic search over it.

How it fits with Hermes Agent

Built as a Hermes Agent skill, with SKILL.md as the agent entry point and a documented pair of Hermes cron jobs that run scripts without an LLM.

How to install brain-vector

These commands are copied from the project's README. Check the repository for the latest steps before you run them.

uv tool install mlx-embeddings --with lancedb
~/.local/share/uv/tools/mlx-embeddings/bin/python scripts/brain_index.py --full

Requirements: macOS on Apple Silicon, uv, and a knowledge base at ~/gdrive/Brain; the first run downloads a model of about 3 GB

Note: It is written around the author's own ~/gdrive/Brain layout of *_sum.md files and requires Apple Silicon.

FAQ

What is brain-vector?

brain-vector is a Hermes Agent skill that indexes summary files from a personal knowledge base into LanceDB. It answers questions through semantic search with metadata filters.

Does brain-vector work with Hermes Agent?

Yes, it is built as a Hermes Agent skill. SKILL.md is the entry point and the question to the brain is the trigger.

What do I need to run brain-vector?

You need an Apple Silicon Mac running macOS, the uv tool and a knowledge base at ~/gdrive/Brain. The first run downloads an embedding model of about 3 GB from HuggingFace.