Our AIE keynote is live! Watch it now

2026-06-26 · Source: Towards AI Newsletter · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Data Science & Analytics · Depth: Advanced, extended

Summary

A keynote presentation from the AI Engineer World's Fair, titled "Turn 10,994 Notes Into Memory," is now available on the AI Engineer YouTube channel. This presentation details a full system developed over 18 months by Louis-François and Paul Iusztin, designed to transform thousands of personal notes from platforms like Obsidian, Readwise, and Notion into persistent agent memory. The system uniquely avoids vector databases or knowledge graphs, relying instead on Markdown, YAML, and folders. It comprises a dynamic Deep Research Agent, which leverages Gemini API for web searches, YouTube video analysis, and GitHub content scraping, and a deterministic Writer Workflow. The workflow employs an evaluator-optimizer loop to produce high-quality, human-like content, while an observability framework using Opic monitors system performance and facilitates AI evaluation. An associated workshop on "Context Engineering in 2026" was also held on June 29, 2026.

Key takeaway

For AI Engineers building agentic systems or content generation pipelines, you can achieve robust, persistent agent memory and high-quality, human-like content without relying on complex vector databases or knowledge graphs. Your focus should shift to leveraging simple file structures like Markdown and YAML for memory, combined with an iterative evaluator-optimizer loop for content refinement. This approach reduces cost and complexity while improving output quality, making it a practical strategy for scalable AI applications.

Key insights

Persistent AI agent memory can be built with simple file structures and iterative refinement, avoiding complex databases.

Principles

• One-shot agents use context, a Research OS builds memory.
• Prioritize simplest solutions; use agents only for dynamic actions.
• Evaluator-optimizer loops refine content and reduce LLM bias.

Method

The system uses a Deep Research Agent for web, YouTube, and GitHub content, compiling to Markdown. A Writer Workflow then iteratively refines content using an LLM, profiles, and an evaluator-optimizer loop.

In practice

• Use Markdown, YAML, folders for persistent agent memory.
• Structure user guidelines to control LLM content generation.
• Implement evaluator-optimizer loops with separate LLM contexts.

Topics

• Agent Memory
• AI Engineering
• LLM Workflows
• Content Generation
• AI Evaluation
• Context Engineering

Best for: AI Architect, AI Engineer, Machine Learning Engineer, AI Student

Related on AIssential

• Agent memory resolved? — A git-like memory framework significantly enhances coding agent performance and context persistence across sessions.
• AI Agents Explained Like You’re an Engineer: A Beginner’s Guide to How Modern AI Actually Works — Modern AI agents operate as complex software systems, integrating multiple steps into a continuous decision-making loop.
• 🥇Top AI Papers of the Week — AI advancements focus on novel computing, efficient memory, robust agents, and specialized models.
• Before You Build an AI Agent, Read This — Effective AI agent development prioritizes foundational workflows and adaptive retrieval over static, single-tool approaches.
• Agent Building Trends [Operator Bonus Episode] — AI agent development is rapidly evolving towards autonomous digital employees and highly personalized "markets of one" software.
• How to Optimize Your AI Coding Agent Context — Effective context management is paramount for optimizing AI coding agent performance and engineer efficiency.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Towards AI Newsletter.