A file-based agent memory framework

Source: Dev.to

Introduction

Hi everyone, we’ve open‑sourced a file‑based agent memory framework called memU. If you find it interesting, a GitHub star ⭐️ would be much appreciated 🙏: GitHub link.

Andrej Karpathy, a member of OpenAI’s founding team, expressed agreement with the view that “RAG is dead.” So we built a memory framework that retrieves knowledge through LLM‑based file reading.

Core ideas

Dual‑mode retrieval: embedding + non‑embedding search

Non‑embedding search is designed to compensate for the structural accuracy limitations of traditional RAG in high‑precision scenarios.

Memory as readable Markdown files, not opaque vectors

At the Category layer, memory is stored as human‑readable Markdown files. This follows the same underlying design philosophy as Anthropic’s skills.md.

Designed to plug directly into real production agents, with fully configurable prompts

For example, an engineering agent can persist its core knowledge as Service_Architecture.md and Incident_Playbooks.md. The model reads these structured files first to establish correct reasoning premises, avoiding drift caused by “similar but incorrect” retrieval results.

Anyone interested is welcome to try it out and explore the code. 😺