Evolving Agents
A director agent that learns team conventions, plans with linked memory, and versions every lesson.
An open-source director agent that learns your team conventions, versioning every lesson into clear patches.
Evolving Agents is a director agent that walks your CLAUDE.md and memory as linked rules to guide coding assistants. It suits teams working with Claude Code or local models who need agents to respect house conventions rather than generic practices.
What it is
The director compiles your memory files into linked rules, injects them before planning, and drafts versioned lessons from your corrections for you to accept or reject. It enforces private task routing to local hardware in code and shows every rule followed or missed in each plan.
Who it's for
It suits teams struggling with coding agents that forget custom definitions and conventions under pressure, or developers requiring strict data privacy for sensitive workloads. It does not suit teams looking for a standalone chat interface or those unwilling to curate structured Markdown memory.
Key features
Linked Rule Memory
Walks your CLAUDE.md and memory statements to hand the exact team rule to the agent before it plans.
Gated Lesson Drafts
Turns your session corrections into proposed versioned patches that require your explicit accept or reject.
Enforced Privacy Routing
Restricts confidential tasks to run only on your cleared local machines or internal servers via hardcoded rules.
Plan Traceability
Lists the exact team rules followed in each plan alongside decisions your current memory does not yet cover.
Up and running in 3 steps
Our honest verdict
What's great
- Carries your team's specific decisions and thresholds directly into the agent planning step.
- Keeps every learned lesson as a readable, comparable, and undoable versioned change.
- Enforces privacy rules in code to ensure confidential code never hits cloud models.
- Integrates natively as a Claude Code plugin with plain Markdown memory staying on your machine.
Where it falls short
- Compiling memory files into linked rules incurs initial token costs of roughly $5 for 60 files.
- Relies heavily on maintaining and structuring your local Markdown memory files accurately.
- Requires using Claude Code and running commands like compile and reflect to capture lessons.
Pricing & license
Free under its open-source license, with no paid hosted option mentioned in the repository; the only cost is running it yourself and compilation token usage.
Open remote AI-training roles
880 open remote AI-training roles, updated November 15, 2026; most pay $20–$175 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
Use it, or have us set it up
FAQ
Does this tool make the underlying LLM inherently smarter?
No. It supplies the context, team decisions, privacy rules, and hard-won methods that frontier models cannot automatically know.
Where is my team's memory stored?
Your memory stays directly on your machine as plain Markdown files.
How are corrections handled?
When you correct an agent, it drafts a lesson that appears in a prompt band allowing you to view the diff and accept or reject it.
