LunaLabs

Machine learning · 2026

Obsidian Mind

Persistent memory for AI coding agents.

Coding agents are capable and amnesiac. Every session starts from zero: the same context re-explained, decisions from three conversations ago lost, and nothing compounding. Obsidian Mind is the fix — a structured vault plus hooks that let an agent read its own history at the start of a session and write back to it at the end.

Status
Open source
Domain
Developer tooling · Open source
Role
Author
Year
2026

Stack

  • Node
  • Obsidian
  • Semantic search
  • Agent hooks
  • Markdown

A knowledge vault that gives AI coding agents memory across sessions — hooks, commands and semantic search, working across three major agent CLIs.

01

Memory as files, not a service

The store is a plain Markdown vault.

That is the whole architectural argument: it is greppable, diffable, version-controlled, editable by a human, portable between tools, and it will outlive any particular agent runtime. Semantic search sits on top for retrieval; nothing sits underneath that a text editor cannot open.

02

Three runtimes, one vault

Hooks and commands are implemented for three separate agent command-line tools, all writing to and reading from the same vault.

Switching agents does not reset the memory, which is the point — the knowledge belongs to the developer, not to the vendor.

03

Writing is the hard half

Retrieval is comparatively easy; getting an agent to write memory that is worth reading later is not.

The vault imposes structure — what a note is for, what belongs in it, how it links to others — because an unstructured log of everything an agent did is indistinguishable from no memory at all. The schema is deliberately small enough that a model follows it without a fine-tune and a human can fix it by hand.

04

Documented for people who are not the author

The project ships architecture notes, contribution guidance and onboarding documentation in four languages.

For a tool whose entire premise is continuity across sessions and across tools, documentation is not an afterthought to the feature set — it is the feature that determines whether anyone else can adopt it.

Pipeline

Session memory

Stage Can reject

  1. 01

    Session start

    Agent asks

  2. 02

    Retrieve

    Relevant notes

    Semantic over Markdown

  3. 03

    Work

    Decisions made

  4. 04

    Distil

    Structured note

    Schema, not free text

  5. 05

    Vault

    Committed memory

Skills exercised

What the build
actually demanded.

Against the corpus · 15 systems

  • Stack breadth 5
  • Design decisions 4
  • Pipeline stages 5
  • Decision gates 1

This system Corpus median

Agent tooling

  • Session lifecycle hooks across three agent runtimes
  • Slash-command surfaces and prompt scaffolding
  • Structured note schemas an agent can write reliably
  • Semantic retrieval over a Markdown corpus

Open source

  • Multi-language documentation and onboarding
  • Contribution guidelines and change management
  • Cross-runtime compatibility testing
  • Architecture documentation for external readers

What it establishes

  • Plain Markdown store

    Greppable, diffable, portable

  • Three agent runtimes

    One shared vault across tools

  • Documented in four languages

    English · Japanese · Chinese · Korean

  • Schema over free text

    Structure is what makes memory worth re-reading