An autonomous agent that writes code, runs it inside a sandbox, then reflects on what it did and what it ought to have done differently. Three kinds of memory, five constitutional principles, and the patience to think before it acts.

Droid8 wants to be a careful colleague. It writes code, runs it inside a Docker sandbox, then sits with the result long enough to learn from it. None of this is loud — the system was made to behave the way a good intern behaves: ask before reaching, learn from the room, keep its own notes.

It carries three kinds of memory — what it has done, what it knows, what tends to work. The first is episodic, a searchable log of attempts. The second is semantic, facts with confidence values that decay if they go untouched. The third is procedural, patterns of code that have earned their keep through use. Each is stored alongside the others; each can be queried by hand.

A constitutional layer sits between thought and action. Five principles — laid out in plain prose, not regulation — decide whether a proposed step is allowed. The agent can argue with them and sometimes wins; more often, it asks for help. This part is the slowest part on purpose.

A program with good manners is a program that has read the room.

The hardware does not matter much. Droid8 runs against a small local model through Ollama, or against larger ones in the cloud when the local one is out of its depth. The router decides which, quietly, and remembers when its choice was the right one.

The source is open. The principles, the prompts, and the patterns are all visible from the outside — you can read them, change them, or write your own.