How to Set Up and Maintain Project Memory

Estimated time: 5 minutes
Difficulty: Beginner

Why you'd use this

Every conversation starts fresh, but a project has standing knowledge: coding standards, things never to change, architectural decisions, the current task list and a log of what was done. Project memory is a small set of markdown files in a .project-ai/ folder that the assistant reads with every request and updates as it works. Commit the folder and the whole team's assistant shares the same context.

Before you start

Permissions required:

• Write access to the workspace root.

You'll need:

• A few lines on the project's conventions and any rules the assistant must respect.

Steps

1. Press / and choose Initialize project memory, or run eLLM: Initialize Project Memory. The folder is created with memory.md, architecture.md, decisions.md, tasks.md and devlog.md, and memory.md opens in the editor.
2. Fill in memory.md: the project's purpose, coding standards, and a "never change" list. Keep it short; it is sent with every request.
3. Work with the assistant as normal. It records decisions in decisions.md, keeps tasks.md current during Plan Mode, and appends a devlog entry when a request changes files.
4. To read or edit memory later, choose Open project memory from the / menu.
5. When the files have grown, choose Compact project memory. The assistant rewrites them shorter, keeping every rule and decision but dropping stale detail, and offers the result as diff cards for you to apply.
6. Commit .project-ai/ to version control, except the audit/, debug/, scratch/ and checkpoints/ subfolders which are local working data.

What you should see

A new project shows "Initialized project memory" with the file names; an existing one shows "Project memory already exists". Helper scripts the assistant writes for itself go under scratch/ and are cleaned up when a build completes. The folder name is configurable with ellmCode.workspaceMemory.

Troubleshooting

• The assistant ignores a standard: check it is in memory.md rather than only in a chat message; only the memory file persists across sessions.
• Memory has become very large: the prompt is fitted to the model's window and a note says which parts were trimmed. Compact the memory, or move long reference material out to normal docs the assistant can read on demand.

Related guides

• How to Resume a Project and Catch Up
• How to Use the Standalone Planning Commands
• How to Check and Manage the Context Window