Open Second Brain is a memory layer for AI agents stored in an Obsidian vault. Everything an agent learns is a Markdown file in a dedicated Brain folder, so you can open, edit, search, and version it like any other note. Agents read and write it through MCP servers, lifecycle hooks, and the o2b CLI, and a rebuildable SQLite index makes it searchable. Your own notes are read, and a note tool writes one only at a path the caller names.
Memory improves through a learning loop. Agents send feedback signals into an inbox, and a deterministic dream pass, with no model inside it, turns repeated signals into preference rules with a confidence level. Active rules are rendered into a note the agent reads at the start of each session, and evidence that a rule was applied or violated confirms, quarantines, or retires it. Rules that nothing uses expire.
Each client reaches the same core through a plugin, an MCP registration, or the CLI. Claude Code, Codex, Hermes Agent, opencode, and Grok Build use plugins. OpenClaw runs it as an in-process plugin. Cursor, Kiro, GitHub Copilot CLI, Gemini CLI, ZCode, and any MCP host connect over MCP. Aider and Pi use the CLI and a context file or skill.
Every memory change takes a snapshot first and can be rolled back, and every note write is attributed to an agent and device in a hash-chained log. Semantic search through an embedding provider, a typed decision model for reranking, and deep relational recall over typed links are optional and off by default.
Features
- Vault-native memory: agent memory stored as Markdown files in an Obsidian vault
- Learning preferences: a deterministic dream pass turns repeated signals into rules with confidence, and retires unused ones
- Reversible changes: snapshots before every change, with rollback
- Attributed writes: each note write is logged by agent and device, with revert planning, a freeze switch, and a hash-chained log
- Search: keyword search with an optional semantic lane, result explanations, and recall profiles
- Session recall: imports agent transcripts already on disk and makes them searchable
- Hygiene scans: finds contested, duplicate, stale, and unused memories and applies only the fixes you pick
- Claim ledger: records entity claims with validity windows and checks agent-stated claims against them
- Corrections: reports a correction's blast radius and retires the affected records
- Knowledge packs and backup: exports chosen subsets or a single-file backup bundle
- Secret custody: encrypted credentials behind a passphrase, referenced by name at use sites
- Many clients: plugins, MCP, or CLI access for Claude Code, Codex, Hermes Agent, OpenClaw, opencode, Cursor, Gemini CLI, Aider, Pi, and more
