Exfoliated Memory: Local AI Knowledge Graph

Purpose and Significance

Exfoliated Memory is a self-maintaining knowledge graph system designed to transform Moltbot from a stateless AI assistant into a persistent, learning entity. Unlike most AI systems that operate without memory between conversations, this project enables automatic fact extraction, contextual understanding, and compound knowledge growth over time. Built on local open-source models, it preserves conversation history without deletion—only superseding outdated information—creating a living repository of learned facts that improve assistant responses indefinitely.

Key Features

• Automatic Fact Extraction — Intelligently identifies and captures factual statements from ongoing conversations without manual intervention
• Entity-Based Storage — Organizes knowledge around entities (people, concepts, relationships) rather than raw text, enabling semantic reasoning
• Weekly Synthesis Summaries — Automatically generates periodic summaries that compound knowledge and reduce information redundancy
• Immutable History — Never deletes facts; instead supersedes them with updated information, preserving an auditable knowledge lineage
• Local-First Architecture — Runs entirely on local models (Qwen 3.4B + Nomic Embed Text via Ollama) with zero cloud dependencies or costs
• SQLite Backend — Lightweight, portable persistence layer requiring no external database infrastructure

Getting Started

Installation requires three prerequisites: a working Moltbot instance, Ollama with the qwen3:4b and nomic-embed-text models, and SQLite3. Clone the repository and run the included setup script to configure the system. Detailed setup instructions are available in SETUP.md within the repository.

Who It's For

This project targets developers and operators building persistent AI assistants, particularly those using Moltbot as their conversational foundation. It's ideal for teams seeking to:

• Build domain-specific knowledge bases that evolve from user interactions
• Create AI systems that demonstrate contextual understanding across time
• Maintain full data sovereignty without relying on external APIs or cloud services
• Reduce hallucination by grounding responses in extracted, verified facts
• Experiment with knowledge graph architectures in a lightweight, self-hosted environment

Architecture Highlights

The system uses a three-component design: conversation monitoring extracts facts in real-time, the entity graph organizes knowledge relationships, and the synthesis engine periodically reviews and consolidates information. Weekly summaries prevent knowledge bloat while preserving important context. By leveraging Qwen's instruction-following capabilities and Nomic's embedding quality, the system achieves high-fidelity knowledge capture without requiring GPU acceleration.

Implementation Notes

Written in Shell for straightforward deployment and integration, Exfoliated Memory prioritizes simplicity over feature complexity. The system is actively maintained (last updated February 2026) and carries an MIT license, enabling both personal and commercial use. With one open issue currently tracked, the project demonstrates active development and community engagement.

Resources

• GitHub Repository — Source code and issue tracking
• SETUP.md — Complete installation and configuration guide
• valeriacross/Moltbot-Start — Primary AI assistant this system integrates with
• DoctorClaw: AI System Diagnostics with Ollama — Local model runtime dependency
• knowledge-graph-basics — Conceptual overview of knowledge graph design patterns

Source: GitHub repository EyeSeeThru/exfoliated-memory, last updated February 2026.