AI

Giving an AI Coding Agent Search and Memory: How the Open myvector Works

myvector is a local CLI that gives Claude Code hybrid search over your projects and a persistent memory, all in one SQLite file and with no external service.

FSM Studio · · 3 min read

Illustration of an open book with a magnifying glass, in terracotta and deep blue

Every new coding-agent session starts from zero: it does not remember last week’s decisions and it searches code by text, not by meaning. myvector is an attempt to fix that. It is an open-source project by the founder of our studio: a personal tool, not an FSM product. Here is what it does and who it suits.

What myvector does

It is a local command-line tool (myvec) that gives Claude Code two things: hybrid search over your projects and a two-level memory (global and per project). Everything lives in one SQLite file (SQLite, sqlite-vec and FTS5).

Hybrid search combines embeddings (vector search) with FTS5 full-text search and ranks the results with RRF. It covers code, docs and tickets, within one project or across all of them:

myvec search "where is VAT calculated"

So the agent can find a place by a question, not only by an exact word.

Memory

Anything worth remembering is saved with myvec remember, for example with --pin. Memory is kept as Obsidian-compatible markdown in a git vault, and those files are the source of truth. The database is derived and can be rebuilt with myvec rebuild.

Secrets and background work

The README describes secret filtering: secret-like files are never indexed and literal secrets are masked. A background daemon, myvec serve, keeps the index fresh with file watchers and offers a local web UI and a JSON API.

Claude Code integration

The tool adds a /myvector skill, session hooks, and tracking of sessions and token usage. It also includes a file-based ticket tracker and project scaffolding for an agent workflow.

Setup

You need Python 3.12, uv, an embedding backend serving embeddinggemma (768 dimensions), either Ollama or llama.cpp llama-server, and macOS or Linux.

git clone https://github.com/Alexshl/myvector && cd myvector && ./install.sh
cd ~/projects/my-project
myvec init --name my-project

install.sh is safe to re-run and installs uv, Ollama with embeddinggemma, the myvec tool, the skill, hooks and the daemon (skip with --no-daemon and --no-hooks). myvec init writes .myvector.toml, registers the project and indexes it in the background.

Limits

  • The full documentation is in Ukrainian; the English README is a short overview.
  • It needs an embedding backend running on your machine.
  • Only macOS and Linux are supported.
  • The license is MIT, so commercial use is allowed.

Who it suits and how to decide

If you run coding agents on large codebases and do not want to send code to an external search service, a local tool is a sensible choice. If you need a supported business solution, see AI automation for business, and read what AI coding tools are. The project is described in the open-source section.

If you are thinking about bringing agents into your own processes, discuss the project with us.

Sources