Markdown-Native Memory Layer for Local AI Agents using Obsidian and Mem0 Principles

[nailulgh]

Hi everyone,

Most local AI coding agents lose context the moment you switch workspaces or close a session. Re-feeding raw conversation logs quickly burns context tokens, while external vector database setups often add unnecessary infrastructure overhead for individual or small-scale workflows.

I built Antigravity Obsidian Memory, an open-source plugin that connects local AI agents directly to an Obsidian vault as a shared, human-readable cognitive memory layer.

Repository: https://github.com/nailulgh/antigravity-obsidian-memory

How it works:

• Router Pattern (< 40 lines): The agent reads a single compact Map-of-Content router first instead of scanning the full vault, cutting context window usage by up to 90%.
• Mem0 Decision Engine: Running /save-memory triggers structured ADD, UPDATE, or NOOP operations to log key decisions and milestone checklists into Markdown notes.
• Cross-Agent Continuity: CLI agents, IDE extensions (Cursor, VS Code), and secondary devices share the exact same context files.
• Plain Markdown & Mobile Ready: Stored as standard notes with Obsidian [[wikilinks]]. Fully readable and editable on mobile via Syncthing or Git without database lock-in.

The project is fully open source under the MIT License. I would appreciate any feedback, thoughts, or ideas on the architecture.