itechmeat/open-second-brain

What it solves

Open Second Brain provides a persistent, user-owned memory layer for AI agents. It eliminates the "black box" nature of agent memory by storing everything as plain Markdown files within an Obsidian vault, allowing users to see, edit, and version their agent's knowledge without relying on proprietary cloud services or hidden state.

How it works

The system integrates with AI agents (primarily Hermes Agent, but also others via the Model Context Protocol - MCP) to read and write knowledge. It uses a deterministic "dream" pass to convert repeated signals into durable rules and retire obsolete ones. Instead of relying on a LLM to manage the memory algorithm, it uses counters and atomic file moves to ensure reliability. It provides a CLI tool (o2b) for management, indexing, and hygiene scans to remove duplicates or contested facts.

Who it’s for

Users of Obsidian who want their AI agents to have a long-term memory that is transparent, editable, and portable across different AI runtimes (such as Claude Code, Cursor, or Aider).

Highlights

• Obsidian-native: Stores all memory as Markdown files, enabling the use of Obsidian's wikilinks, backlinks, and graph view.
• Cross-agent compatibility: A single memory vault can be shared across multiple AI runtimes via MCP.
• Deterministic Learning: Uses a non-LLM algorithm for the "dream" pass to avoid hallucinations in the memory layer.
• User Control: Includes CLI tools for snapshots, rollbacks, and hygiene scans to clean up stale or duplicate knowledge.
• Provenanced Knowledge: Tracks where information came from, distinguishing between operator-stated rules and machine-derived facts.