memU: Enabling Long-Term Memory for AI Coding Agents

Why it is needed

AI coding tools without a memory system treat each conversation as a first meeting, requiring repeated description of project architecture, coding style, and past decisions. In collaborative projects, agents cannot acknowledge prior work.

Solution

memU organizes information into a structured Markdown file tree. An agent loads only the required files based on a given path, avoiding inclusion of the entire history in each prompt.

Usage

Python library

from memu import Service
await service.memorize(resource_url="workspace/meeting-notes.md", modality="document")
context = await service.retrieve(
    queries=[{"role": "user", "content": {"text": "关于这个用户的什么偏好需要知道？"}}]
)

CLI tool

npx memu-cli memorize-workspace ./workspace
npx memu-cli retrieve-workspace "what do I need to know?"

Memory content

memU supports dialogue logs, code files, documents, images, audio, video, web links, and tool‑call records. These are compiled into three core Markdown files:

INDEX.md – index for quick navigation

MEMORY.md – stored memory content

SKILL.md – description of the agent’s skills

The agent traverses the file tree and loads only needed pieces, preserving long‑term memory without wasting tokens.

One‑sentence summary

memU provides long‑term memory for AI agents: define project structure and preferences once, and the agent automatically loads them on subsequent runs.

Stars: 13.9k+ Language: Python License: Apache 2.0

Project URL: https://github.com/NevaMind-AI/memU