The cure for AI amnesia is an Obsidian notebook

What it does

obsidian-mind is a structured Obsidian vault that acts as a long-term memory bank for AI coding agents. Instead of starting each session with a blank slate, Claude Code (and experimentally Codex CLI and Gemini CLI) loads your goals, active projects, recent decisions, and team context from markdown notes. The agent then files new decisions, meeting notes, and brag-doc entries back into the vault graph, linking them to people and projects automatically.

The interesting bit

The project draws a hard line between deterministic glue code and agent judgment: TypeScript hooks handle context injection, validation, and indexing, while the agent decides what to write and where to link it. This keeps the agent’s context window lean—only about 2K tokens of base context are injected at session start, with semantic search via QMD fetched on demand—while still giving the agent a browsable, git-tracked knowledge graph.

Key highlights

• Lifecycle hooks (SessionStart, UserPromptSubmit, PostToolUse, PreCompact, Stop) classify your chatter and route it into the right notes without manual prompting
• Vault-first memory: durable knowledge lives in brain/ markdown files, not in the agent’s sandboxed memory file, so it survives machine swaps and is readable in Obsidian
• Optional QMD semantic search plugs in as an MCP server, letting the agent query “what did we decide about caching” even if the note is titled “Redis Migration ADR”
• Token-efficient tiered loading: always-on context is limited to your North Star, active work, and file listing; the full vault is never dumped into the prompt
• Built primarily for Claude Code, with working but less complete hooks for Codex CLI and Gemini CLI

Caveats

• QMD’s full semantic search pulls down roughly 1.6 GB of models on first use, though lighter BM25 or vector-only modes are available
• Hook scripts rely on Node’s --experimental-strip-types flag, which the README notes may require a one-line fix if Node retires it in a future release
• Codex and Gemini support is explicitly “hooks + commands” level, not the “full support” given to Claude Code

Verdict

Worth a look if you live in Claude Code and want your project context to compound across sessions. Skip it if you expect a plug-and-play app—this is a vault template, a workflow, and a set of hook scripts that demands buy-in to its note-taking religion.