memini-ai

Local-first memory for AI coding agents. Three MCP tools: remember, recall, orient. Postgres + pgvector for storage, BGE-M3 embeddings on CPU, hybrid vector + full-text search.

Quick start

uv tool install memini-ai --index https://download.pytorch.org/whl/cpu --index-strategy unsafe-best-match
memini-ai db up            # pgvector/pgvector:pg18 on 127.0.0.1:5555
memini-ai migrate          # optional; serve applies migrations on first connection
memini-ai warm             # one-time ~2.2 GB BGE-M3 download, then embeds one string

The --index flags install CPU torch instead of the default CUDA build. Run warm before wiring up a client: the first remember or recall would otherwise download BGE-M3 (about 2.2 GB) inside a tool call and time out until it finishes.

Then, from the project you want memory for:

cd ~/Projects/my-repo
memini-ai init --client claude-code --project my-repo

Restart the client and call orient.

Docs

• Getting started
• The three tools
• Configuration
• Clients and the init-memini skill
• Session ingest
• Architecture
• Changelog

License

MIT. See LICENSE.