A local-first agentic framework acting as a personal operating system, leveraging file intelligence, event-driven workflow automation, and LLMs for cross-modal task execution and multi-platform interaction.
Second Brain is a local-first agentic framework designed to serve as a personal operating system for digital assets. It continuously monitors designated directories, automatically building semantic indexes across all modalities—text, PDF, DOCX, PPTX, spreadsheets, archives, images (OCR), audio, and video—offering four retrieval modes: lexical search, semantic search, hybrid ranking, and direct SQL queries.
For task orchestration, Second Brain implements an internal pub/sub event bus supporting cron scheduling, chained background execution, approval workflows, and proactive notifications. Tasks are managed via DAG dependency graphs with pause, retry, and timeout recovery capabilities. The Agent maintains persistent memory (memory.md) and full conversation history for cross-session context continuity.
A standout capability is self-extension: the Agent can dynamically create, edit, and delete plugins (services, tasks, tools) in a sandbox via build_plugin, with hot-loading requiring no restart. For complex tasks, the primary Agent can delegate to sub-agents through ask_subagent or schedule_subagent.
Interaction is delivered through a Telegram Bot (slash commands, file push, approval prompts, background task notifications) and a terminal REPL, with modular frontend code extensible to Discord and other platforms. Web search integrates Brave Search and DuckDuckGo, automatically supplementing local knowledge with public web information when needed.
Architecturally, all frontend operations are dispatched through a unified ConversationState + ConversationRuntime state machine. The system prompt is dynamically reconstructed per message to reflect the latest tools, services, tasks, and file states. All data—task queues, conversation history, embedding indexes, file metadata—resides in a single SQLite database, achieving zero external dependency for local deployment. The LLM backend supports OpenAI API and compatible endpoints like LM Studio, with flexible multi-model and multi-agent configuration through llm_profiles and agent_profiles.
Unconfirmed information: Open-source license not explicitly stated in README or repository root; no independent website or documentation site found; no associated papers or Hugging Face models/datasets discovered; minimum Python version requirement unclear; no performance benchmarks or evaluation data provided.