Your AI coding agent forgets your project every session. Here is what I built about it

Disclosure up front: I build Codelux, the tool in this post. There is a free plan and paid plans. I'll try to keep this useful even if you never install it.

The problem you already pay for

If you use a coding agent every day, you know the morning routine. You open a new session and the agent knows nothing. It re-reads the same files to rebuild the same picture of the codebase it had yesterday. You explain the architecture again. You remind it what was decided and why.

That costs you twice:

• Tokens. Whole files go into the context window to find one function, and the whole conversation is re-sent on every turn.
• Decisions. Why the code is built that way, what was tried and dropped: it lived in a chat that is gone. The agent may happily re-introduce the thing you removed last week.

And if two people, or two agents, work on the same repository, they share nothing they learned. Each one rebuilds the same context on its own.

What helps, with or without any tool

Before the product part, the habits that made the biggest difference for me:

1. Locate first, then read a window. Grep or a symbol index to find the function, then read 30-60 lines around it. Reading a 400-line file to use 12 lines of it is the most common waste in a long session.
2. A new session per task, with a short handoff. Goal, decisions taken and why, files touched, what is left. Ten lines in, instead of the whole history re-sent every turn.
3. Split notes by how they age. Decisions and their reasons stay true, so write them down. Anything structural (where X lives, what calls what) rots quietly, so look it up in the code at read time instead of writing it into a file that will be wrong in a month.

You can do all three with plain markdown files and grep. I did, for a while. Then I automated it.

The 2-minute demo

What you see, all recorded on the real app:

• Indexing in seconds. A 120-file Go project (the open-source gin framework) indexed locally in about 3 seconds: files, symbols and call flows go into a local database, and it stays up to date as you edit.
• Notes and decisions that carry over. While the agent works it leaves notes and decisions in the project. The next morning a new Claude Code session calls code_memory and starts from yesterday instead of from a tour of the repo.
• Lookups instead of whole files. code_query answers "where is X, who calls it" with file and line, and file reads come back in a compressed form: up to 90% fewer tokens on a single read.
• The Team plan. Two developers on the same project: a decision one writes shows up on the other's machine in under a second, end-to-end encrypted, next to a team chat.

How it plugs in

Codelux is a small local daemon. Your agent talks to it over MCP on localhost (HTTP and a CLI also work), so it fits whatever you already use: Claude Code, Cursor, Codex, VS Code, Windsurf, JetBrains, Zed or any MCP client. For Claude Code there is a plugin with a setup skill:

claude plugin marketplace add codeluxai/codelux
claude plugin install codelux@codelux

Your code, the index and the memory stay on your machine. Nothing is uploaded; AI calls go only to the provider you choose.

What it does not do

To keep expectations honest:

• It is not another agent and it does not write code for you. It is the space your agents work in.
• On a small repo, in a single short session, a good grep is often just as cheap. The benefit shows up across sessions and across people, when the same understanding would otherwise be rebuilt again and again.
• The daemon is not open source. The docs, the plugin and the issue tracker are on GitHub.

Try it

The Free plan is free forever for one project per machine, no credit card: codelux.ai/download

Docs: codelux.ai/docs · Issues and plugin: github.com/codeluxai/codelux

I'd genuinely like to hear how you handle agent memory today, especially if your answer is "markdown files and discipline". That is where I started.