Kanjironomish/MEMORY_AUGUMENTED_AI_ENVIRNMENT

Hackathon AI Environment

This project turns the shared design into a runnable prototype:

User Query
   -> Vector Encoding
   -> Multi-Agent Layer
   -> Shared Knowledge Space
   -> Decision Fusion Engine
   -> Q-Learning Controller
   -> Deterministic Reward Node
   -> Memory Update with Decay
   -> Next State

It includes four agents:

• —food
• —business
• —research
• —memory

The environment is designed for hackathon demos where you want:

• —multi-agent routing
• —deterministic, judge-friendly rewards
• —a real Q-table update loop
• —memory recall across sequential queries
• —a clean CLI for training and evaluation

Project Layout

• —hackathon_ai_env/agents.py: specialist agents and query-domain encoding
• —hackathon_ai_env/fusion.py: decision fusion and action generation
• —hackathon_ai_env/q_learning.py: epsilon-greedy action selection and Q updates
• —hackathon_ai_env/reward.py: deterministic reward scoring
• —hackathon_ai_env/memory.py: shared memory with decay and reinforcement
• —hackathon_ai_env/environment.py: end-to-end environment loop
• —hackathon_ai_env/scenarios.py: sample benchmark tasks
• —main.py: CLI entrypoint

Run It

Train on the built-in scenarios:

python3 main.py train --episodes 40

Train and evaluate the learned policy:

python3 main.py eval --episodes 40

Train, warm shared memory, and ask a custom question:

python3 main.py ask "What pricing model should we use for a student startup?" --episodes 40

Launch the browser dashboard:

python3 main.py web --port 8000 --episodes 40

Then open http://127.0.0.1:8000.

Deploy On Hugging Face

This repo is now set up for a Hugging Face Docker Space.

1. 1.Create a new Space and choose the Docker SDK.
2. 2.Push this entire repository to that Space.
3. 3.Hugging Face will build the included Dockerfile and start the dashboard on port 7860.

The container starts the app with:

python3 main.py web --host 0.0.0.0 --port ${PORT:-7860} --episodes ${EPISODES:-40}

Optional runtime variables you can set in the Space settings:

• —EPISODES: default training episodes for the dashboard, defaults to 40
• —PORT: app port, defaults to 7860
• —HOST: bind address, defaults to 0.0.0.0
• —APP_STATE_PATH: persisted dashboard state file, defaults to /data/hackathon_ai_env_state.json in the Docker image

Notes for deployment:

• —Hugging Face Docker Spaces expect a single exposed app port; this repo uses 7860.
• —The dashboard now autosaves the Q-table, memory bank, training summaries, and latest feedback to APP_STATE_PATH, and reloads them on startup.
• —To keep that state across Hugging Face restarts, enable persistent storage for the Space so /data survives restarts.
• —Local browser URLs such as 127.0.0.1 are only for local development; Hugging Face will provide the public Space URL after deployment.

What The Q-Learning Action Means

Each action is a compact policy choice:

• —which agent to trust
• —whether to use shared memory
• —what confidence threshold must be met before accepting that agent

If the selected agent does not clear the threshold, the environment falls back to the fusion winner.

Reward Design

The reward node is deterministic and reproducible. It scores:

• —domain accuracy
• —keyword coverage
• —memory usage quality
• —confidence alignment

This keeps the system hackathon-friendly because the grader does not depend on a stochastic LLM reward.

Tests

Run the lightweight test suite with:

python3 -m unittest discover -s tests