← All Thoughts
Knowledge ManagementAI AgentsObsidianPARA

MindPalace: A Regimented Multi-Agent Framework

Originally the README for MindPalace, an open-source framework I built and maintain on GitHub.

MindPalace is an infrastructure blueprint for building and maintaining a “Digital Brain” using Obsidian and LLM agents. It combines the structured rigor of PARA/Life Horizons with the emergent knowledge synthesis of the Karpathy LLM-Wiki pattern.

The Problem

Most Obsidian vaults turn into “digital graveyards.” We collect links, screenshots, and PDFs, but we rarely synthesize them into actual wisdom.

Our Solution: The Persona-Based Approach

Instead of a single AI assistant, we orchestrate three distinct “Persona” agents to maintain your vault:

  1. The Operator (Hygiene): Handles the grunt work—archiving logs, linting links, and enforcing folder schemas.
  2. The Curator (Synthesis): Focuses on ingestion, MOC (Map of Content) generation, and cross-linking concepts.
  3. The Researcher (Deep Ideation): Deep dives into topics, comparative analysis, and high-level synthesis.

Features

  • Horizon-Based Organization: Transitions notes from Active (immediate focus) to Incubating (development) to Evergreen (synthesized wisdom).
  • Provenance-First Ingestion: Every note tracks its source (extracted/inferred/ambiguous) to prevent hallucination drift.
  • Agent Hygiene: Automated archival and maintenance scripts to keep your vault performant without manual intervention.
  • Seamless Web Capture: Integration with the official Obsidian Web Clipper using standardized template triggers for automatic ingestion into 00_Inbox/_raw.
  • BYOM (Bring Your Own Model): Every workflow is decoupled from the LLM provider. Whether you self-host (Llama, DeepSeek) or use premium APIs (Gemini, Claude), the framework works the same.

How to Get Started

  1. Clone this repo.
  2. Bring Your Own Agent (BYOA): Point your preferred agent (Cursor, Aider, Hermes, Windsurf) to this directory.
  3. Initialize: Copy the templates/ into your Obsidian vault and configure the scripts/ in your agent’s task scheduler (or cron).
  4. Configure Personas: Define your “Operator,” “Curator,” and “Researcher” personalities in your agent’s config file (e.g., .aider.conf.yml or hermes-agent.toml) using the provided workflows/.
  5. Ingest: Start dropping raw captures into your 00_Inbox/_raw folder and trigger your Curator persona.

Attribution

This framework is built upon the foundational Karpathy LLM-Wiki pattern.

← Back to All Thoughts