abmind

This Skill addresses the lack of long-term memory in AI agents. In practical development, developers often need agents to remember cross-session context, user preferences, and historical interactions, but most solutions rely on vector databases and lack multilingual support. abmind implements a 4-layer retrieval system (FTS5 full-text search + trigram fuzzy matching + embeddings + merged summaries) based on SQLite, and includes a built-in nightly sleep maintenance mechanism that automatically compresses, merges, and cleans memories to prevent context bloat.

Usage is straightforward: simply install the abmind package in your project, then call the recordMessage() API to store messages. Natural language queries will then trigger the 4-layer retrieval. The system automatically classifies, scores, and encrypts memories (e.g., credentials are encrypted using AES-256-GCM). During the nightly sleep phase, maintenance tasks such as summary extraction, Darwinian selection, and emotional arc analysis are performed.

It is particularly well-suited for multilingual AI agents that need to run long-term (24/7), especially those handling agglutinative languages like Japanese and Turkish. For scenarios requiring persistent memory—such as building intelligent customer service, personal assistants, or game NPCs—abmind provides a lightweight yet complete solution without relying on external vector databases or cloud services.

When deploying, ensure Node.js version ≥22 and periodically check the SQLite database size. Note that it primarily focuses on memory storage and retrieval and does not include dialogue generation logic; it must be used in conjunction with an LLM interface. Additionally, its embedding layer uses a local model by default; for higher accuracy, it can be replaced with an external embedding API.