What is project memory for AI coding agents?
Project memory is durable, reviewable context about a codebase: how it is shaped, which decisions are in force, which conventions apply, and where the last person left off. It lives with the repository so the next engineer and the next agent session start from the same facts. It is not a chat log, not a vector dump of messages, and not a giant rules file loaded on every turn.
Session memory dies when the tab closes
A chat holds what this session just said. Close it, and that thread is gone. The next session does not inherit the decision you made an hour ago unless something outside the chat recorded it.
Rules files tell the agent how to behave
CLAUDE.md and AGENTS.md tell the agent how to behave: commands, style, guardrails. They rot when they also try to store everything the team has learned. A file that loads on every turn is the wrong place for the whole history of the project.
Code search knows the tree, not the reason
Code search knows what the tree looks like. It does not know why a constraint exists, which alternative was rejected, or what the last person handed off.
Project memory is the accumulating record
Project memory is the accumulating record, grounded in the code it describes. It stays next to that code so a teammate can read it and an agent can retrieve it when the task needs it.
mex is one implementation of project memory. Open source is the git-native CLI. Hosted is early access. Install the CLI · OSS vs hosted.