Skip to main content

​
What ICRL Is

ICRL (In-Context Reinforcement Learning) is a trajectory-learning framework for LLM agents. It works by:
  1. Running tasks in an environment.
  2. Storing successful trajectories.
  3. Retrieving similar prior steps during future runs.
  4. Curating low-utility trajectories over time.

​
What You Get

  • Python package: icrl
  • TypeScript package: icrl
  • Python CLI: icrl (tool-calling coding assistant)
  • TypeScript web demo: Next.js + Convex example

​
Package Scope

  • Python package focuses on:
    • ReAct loop (Agent, ReActLoop)
    • FAISS-backed TrajectoryDatabase
    • Built-in providers (LiteLLMProvider, AnthropicVertexProvider)
    • CLI and database utilities
  • TypeScript package focuses on:
    • Same algorithmic abstractions (Agent, TrajectoryDatabase, TrajectoryRetriever, CurationManager)
    • Pluggable storage (StorageAdapter), including FileSystemAdapter
    • Built-in providers (OpenAIProvider, AnthropicProvider, AnthropicVertexProvider)

​
How The Algorithm Runs

  1. reset(goal) on environment.
  2. Generate plan using retrieved examples.
  3. Repeat reasoning/action/observation steps.
  4. If successful in training mode, store trajectory.
  5. Update retrieval feedback and run curation periodically.

​
Next Steps

  • Start with /installation
  • Build first run at /quickstart
  • Read algorithm details at /core-concepts/icrl-algorithm