eugeniughelbur/obsidian-second-brain

Worth a look if you keep notes in Obsidian and already run an agent CLI daily; the value compounds only if you let it write into the vault consistently. This is a skill/command pack, not a standalone memory service: the vault is the store, and the model does the filing.

The problem

Every agent session starts empty, so you re-explain your projects, recent decisions and preferences before asking anything useful, and the conversation's output ends up pasted into notes by hand. With this installed, a command like /obsidian-save takes a finished meeting or chat, extracts decisions, people, tasks and ideas, and files each into the right note. A recorded memo or whiteboard photo goes through /obsidian-ingest, which transcribes or reads it and distributes extracted items across entity pages, task boards and the daily note. The result is that later sessions can query accumulated notes instead of receiving a fresh briefing.

What's different

The mechanism matters more than the command count. Notes are ordinary Markdown files in the vault, so the memory outlives the CLI you happen to use and can be read, edited and backed up with normal tools. The project describes extending Karpathy's LLM Wiki pattern by rewriting existing pages when facts change, not only appending new ones, and offers /obsidian-reconcile for contradictions; that is the author's description, not an independent evaluation of how well the rewrite behaves on messy vaults. "Crew" is a separate, Obsidian-optional feature: Claude Code, Codex, Cursor and Grok Bot record jobs in one plain-folder logbook so one agent can see what another finished and leave a lesson for review. A 49-command set (48 in builds without the Google Calendar MCP) spans search, thinking prompts, research and scheduled background agents, which is a lot of surface to learn; the trade-off is that consistency comes from your note habits, and automatic rewriting means a mistake can propagate into existing notes unless you review diffs or keep version history.

Who should try it

Good fit: someone with a substantial Obsidian vault who uses Claude Code or another supported agent CLI several times a week, wants decisions and meeting output captured without manual filing, and is comfortable reviewing how their notes get edited. The multi-agent logbook is useful if you switch among Claude Code, Codex, Cursor and Grok Bot and lose track of what each did. Poor fit: people without an existing vault or note discipline, anyone who only wants a chat client, and teams needing a shared memory service with access control rather than files in a personal vault. A lighter path exists: the README publishes the note rule as a pasteable block that installs nothing and works without the project, which suits people who want the convention but not the 48-command toolkit.

Requirements and limitations

The primary requirement is a supported agent environment and a vault it can write to; install paths differ per build (CLAUDE.md, AGENTS.md, GEMINI.md, .agents/skills/, .pi/ or a workflow skill file), and the README publishes a CI-generated conformance table showing all builds passing a basic file-tree and script-path check. That check verifies packaging, not memory quality. Word and OpenDocument ingestion is an extra chain: it converts a private copy to a temporary PDF and needs Python 3.10+ plus LibreOffice 26.2+ installed separately, with comments and revision markup excluded. Updates are manual, and on Windows require running an extra script to refresh copied command files, so this is ongoing upkeep rather than install-and-forget. Search quality is claimed with a reproducible 300-note synthetic benchmark, which is testable but not a substitute for trying your own vault.

HotRepo's take

HotRepo has only one dated observation: 4,925 stars and +35 on the daily board on 2026-10-26, with no earlier history, so persistence of interest cannot be judged from that sample. Treat the stars as a snapshot. The concrete thing to verify before adopting is how the rewriting and reconciliation commands behave on your own notes: check the command reference and read the benchmark setup, then decide whether you want an agent editing existing pages at all. Command reference · Repository · Retrieval benchmark · Rule-only block