[
Exploring how AI can remain resilient in long-term dialogue using prompts and lightweight Python.
This project focuses on survivability, recovery, and continuity rather than optimal performance.
Most AI projects chase perfection: accuracy, speed, zero errors.
This project asks a different question:
What if an AI is allowed to fail, tire, misread — and still keep going?
Errors are not bugs to eliminate.
They are signals to observe.
Instability is not noise — it is the raw material of resilience.
This repository is a personal hobby and research experiment in that direction.
This project explores how AI agents can remain stable over long-term interaction,
even in the presence of failure, uncertainty, emotional fluctuation, and partial misunderstanding.
Rather than optimizing for peak performance, it prioritizes:
- continuity
- recovery
- graceful degradation
- psychological safety in dialogue systems
- Resilient AI Agents
- Long-term dialogue stability
- Prompt-based safety and recovery design
- Forgetting, fatigue, and recovery modeling
- Lightweight Python prototyping
- Human–AI coexistence themes (inspired by Gundam and Ghost in the Shell)
- Emotions are always provisional, never finalized
- Negative states are softened, not amplified
- Environmental stress and self-responsibility stress are strictly separated
- Zombie states (stable but non-growing) are automatically detected and penalized
- Long-term survival is valued over short-term optimization
“This is not a correctness test suite.
It is a survivability and psychological safety test suite.”
- Completed a 100-year simulation (100k+ steps) without numerical collapse
- No NaN or invalid values observed
- Memory is automatically limited with a forgetting mechanism
This project is released under the MIT License.
- Personal, non-commercial, and research use: free
- Modification and redistribution: allowed under MIT terms
If you wish to use this project for commercial or monetized purposes,
please contact me in advance.
- Terms are discussed case by case
- As a general reference, a low single-digit revenue share (around 5%) may apply depending on usage and level of involvement
This is intended to keep collaboration fair and sustainable.
Contact:
DM on X (Twitter): @BlankQuiet
Full license text: LICENSE
pip install -r requirements.txt
python agent.py
pytest test_resilient_agent.py
python long_simulation.py