gbase
GBase — Recursive Self-Improvement Agent Framework. Memory, evolution, quality gates, identity system, and 40+ auto-registered tools.
A Python framework for building AI agents that remember past sessions, can rewrite their own instructions to improve over time, and come loaded with 40+ tools, compatible with any OpenAI-format API including local models.
GBase is a Python framework for building AI agents: software programs that use a large language model to think through tasks and take actions automatically. Unlike simpler agent setups where the AI handles a request and immediately forgets the context, GBase gives agents a persistent memory system so they can recall things from earlier sessions, not just the current conversation.
The framework's headline feature is what the README calls recursive self-improvement. When triggered by calling a specific method in code, the agent analyzes its own past outputs, identifies patterns where it performed poorly, and rewrites its own instructions before the next run. This is not a process that runs automatically on its own; it is something a developer or the agent itself deliberately kicks off.
Other built-in components include a quality gate system where separate agent roles review each other's work before output is accepted, a scheduler for setting up recurring tasks without an external scheduling tool, and an experience engine that distills past interactions into structured knowledge the agent can query later. Over 40 tools come pre-loaded, and new ones can be added by writing a Python function with a decorator, which makes the function automatically available to the language model without extra wiring.
The framework works with any API following the OpenAI format, including hosted services and local setups. Configuration is done through an environment file where you set your API key and preferences. Multiple agent personas can be defined and swapped, and the framework includes a sub-agent mode for running specialized roles like code review or frontend work on separate local ports.
The author describes this as one part of a broader set of projects, with separate tools for multi-agent coordination and a virtual environment where agents can operate with persistent identities.
Where it fits
- Build an AI agent that remembers earlier conversations and recalls past experiences when handling new requests in later sessions.
- Add a custom tool to your agent by writing a Python function with a decorator so the language model can call it automatically without extra wiring.
- Trigger the self-improvement method to have your agent analyze its past outputs, spot weak patterns, and rewrite its own instructions for better future runs.
- Schedule a recurring automated task directly inside GBase without needing an external scheduling tool like cron or Celery.