henrydaum/second-brain

Implements a **self-extending plugin architecture** where the LLM agent can create new tools, tasks, and services at runtime by writing Python to a sandbox directory, enabling capability expansion without restarts. The four-stage pipeline (Services → Parsers → Orchestrator → Agent) uses filesystem watching with SQLite task queuing, automatic dependency resolution, and modality-aware routing to continuously index multimodal files—then exposes everything through a hybrid search tool that fuses BM25 full-text indexing with vector similarity via Reciprocal Rank Fusion. Built on Flet for the GUI with OpenAI-compatible LLM backends, SentenceTransformer + CLIP embeddings, and native Windows OCR.

432 stars.