Motivation
AI agents in production break down as conversations grow long. They lose track of key facts, get confused by contradictions across sessions, and eventually exceed their context window. This environment simulates that exact problem.
Compressing "User mentioned their name is Venky at 8am, then at 9am they said they love coffee"
into "User: Venky, loves coffee."
Technical Core
| Parameter | Specification |
|---|---|
| Environment | OpenEnv RL Wrapper |
| Engine | FastAPI / Pydantic v2 |
| Complexity | Easy, Medium, Hard tiers |
| Observation | Full text history + Metadata |
Reward Intelligence
| Criterion | Weight | Mechanism |
|---|---|---|
| Fact Retention | 0.5 | Semantic similarity check |
| Conciseness | 0.2 | Token reduction ratio |
| Contradiction Resol. | 0.2 | Logical consistency score |
| Hallucination | -0.3 | Penalty for ungrounded info |
Implementation
# Install dependencies
pip install openenv fastapi pydantic
# Run the compression engine
uvicorn app.main:app --port 7860Designed for the Meta x Scaler Hackathon 2026.