LEARN / SELF-EVOLUTION
The weights stay fixed.
The system gets better.
A stronger model is only one kind of progress. Memory, skills, tools and orchestration can improve too — under evidence, and reversibly.
What it learns.
Memory, by scope
What holds for you goes into user memory; what holds for one folder stays with that project. Facts can expire, and recall has an admission bar so stale notes stop being used.
Skills
Procedures distilled from successful work. Yours are written in the same format as the built-in ones and pass the same quality checks.
Tools
New ways to act — plugins, connectors, small tools — introduced through review and bounded by your permissions.
Change is not the same as improvement.
Observation, proposal, comparison, decision, application and rollback are separate stages. A candidate can be generated automatically; whether it is an improvement is a question of evidence. It is tested against comparable work in an isolated evaluation host and must pass a regression gate. Every step is recorded in a ledger you can replay, and any change can be undone.
Two switches govern it: background tidying and idle-time skill induction are on by default and only produce candidates; automatic promotion is off, so a new skill waits for your approval. Read more on the evolution page.
EvoForge as a coding agent
An original Rust kernel, syntax-aware edits, tests that must fail first, and teams of specialist agents.
Read04 / ABOUTAbout EvoForge
Who makes it, what it is for, and the principles it is built on.
Read01 / OVERVIEWWhat is EvoForge?
The one-paragraph answer, the editions, the platforms, and how it differs from other projects with the same name.
ReadLet it learn your work.
Windows and macOS. Bring your own model key or a local model.