Claude Code memory, with your judgment in it.
Claude Code can be very good at the work and still ask you the same small questions every time. Usual carries the decisions you have already made into the next build.
Usual is an open-source local memory skill for Claude Code. It finds question-and-answer moments in your local Claude and Codex conversations, keeps their context, and gives your current agent relevant evidence when a similar choice comes up.
What it remembers
Usual is interested in the moments when you chose: SQLite instead of another service, a native browser API instead of a dependency, a local demo instead of an account system. It keeps the question, the alternatives, your answer, the reason you gave, and where the evidence came from.
That is different from writing a permanent rule for every possible future project. Context still matters. A preference can have an exception, and a new choice can be reviewed before it becomes evidence.
How it fits into Claude Code
Install the skill
Tell Claude Code: Learn Usual: https://tryusual.com/learn.md. The installer uses Python 3.11+ and has no extra model account or pip dependency.
Let it learn your calls
Use /usual to choose recent history, all local history, or a clean start. Usual never modifies your original conversations.
Build with the context available
When a similar decision appears, the skill shows the evidence to Claude Code. The agent can make the routine call, ask you to check in, or escalate when the evidence is weak.
Usual is not a magic memory layer
It does not promise that an agent knows what you would choose. It gives the agent better examples and keeps the reasoning visible. New predictions remain separate until you explicitly accept or correct them.
It also does not make Claude Code itself run offline. Usual's own mining, storage, retrieval, and review run locally; selected excerpts are still processed by your existing coding model under its normal provider policies.
Try it
Start with the installation guide, read what stays local, or see the worked reading-list example.