agent-second-brain

Accepts voice notes via Telegram, transcribes them using OpenAI's Whisper API, then parses the transcription through Claude to extract entities, relationships, and semantic meaning. The system converts unstructured audio into structured knowledge graph nodes with metadata (source, timestamp, confidence scores). Integration with Telegram Bot API enables real-time voice message capture and processing through OpenClaw orchestration layer.

Implements the Ebbinghaus forgetting curve algorithm to score knowledge items based on review frequency and time intervals. Each note tracks review history, calculates decay probability using exponential decay functions, and assigns a freshness score (0-100). The system prioritizes items approaching the forgetting threshold for review, enabling evidence-based spaced repetition without manual scheduling. Decay calculations run on-demand during vault health scoring cycles.

Exports knowledge base to Obsidian-compatible markdown format with frontmatter metadata (tags, relationships, decay scores, review dates). Maintains bidirectional compatibility: notes created in agent-second-brain can be edited in Obsidian, and changes sync back. Uses standard markdown + YAML frontmatter, enabling interoperability with other tools. Supports Obsidian plugins like graph view, backlinks, and dataview.

Builds a directed graph of knowledge items by extracting entity mentions and relationships from notes using Claude's semantic understanding. Nodes represent concepts/entities; edges represent relationships (e.g., 'mentions', 'contradicts', 'builds-on'). The system infers implicit relationships by analyzing note content and cross-referencing existing nodes, enabling discovery of unexpected connections. Graph is stored as adjacency lists with edge metadata (relationship type, confidence, source note).

Calculates a composite health score (0-100) for the knowledge vault by analyzing multiple dimensions: note coverage (breadth of topics), depth (detail per topic), decay distribution (how many notes are at risk of being forgotten), graph connectivity (orphaned vs well-connected nodes), and consistency (contradictions or duplicate knowledge). Runs periodic scans and generates diagnostic reports highlighting weak areas. Score is weighted and configurable per user priorities.

Generates a daily report summarizing vault activity, highlighting notes due for review (based on decay scores), new connections discovered in the knowledge graph, and vault health changes. Uses Claude to create natural-language summaries of key insights rather than raw data dumps. Reports are formatted as markdown and delivered via Telegram, with optional export to email or Obsidian. Scheduling uses cron-like patterns (configurable daily time).

Automatically creates tasks in Todoist from voice notes, extracting action items using Claude's semantic understanding. Each task includes context from the original note, related notes from the knowledge graph, and decay-based priority (high priority for notes approaching forgetting threshold). Tasks are tagged with source note ID and vault health indicators. Integration uses Todoist API with OAuth authentication. Bidirectional sync allows task completion to update note review history.

Maintains persistent state across sessions by storing note metadata, review history, decay scores, and graph structure in a local database (likely SQLite or JSON files). Each note record includes creation timestamp, review timestamps (array), decay score, last updated, and relationships. State is loaded on startup and persisted after each operation. Handles concurrent access via file locking or transaction management. Enables recovery from crashes and audit trails of knowledge evolution.

Uses OpenClaw (a framework for composing LLM agent workflows) to orchestrate multi-step processes: voice transcription → semantic parsing → entity extraction → graph updates → task creation → report generation. Each step is a declarative node in a DAG, with Claude as the reasoning engine. OpenClaw handles state passing between steps, error handling, and retry logic. Enables complex workflows without custom orchestration code.

Uses Claude's language understanding to parse unstructured note content and extract entities (people, concepts, dates, locations), relationships, and semantic tags. Claude performs zero-shot extraction without training data, understanding context and domain-specific terminology. Extraction results include confidence scores and explanations. Supports custom extraction schemas via prompt engineering. Results feed into knowledge graph construction and task creation.

Implements a Telegram bot using the Telegram Bot API that listens for incoming voice messages, text messages, and commands. Voice messages are forwarded to Whisper for transcription; text is processed directly. Bot maintains conversation state per user (e.g., awaiting confirmation, in edit mode). Supports commands like /review, /health, /export. Uses polling or webhooks for message delivery. Handles rate limiting and error recovery gracefully.