Memevolution

Inspiration

Generative AI can create almost unlimited content, but most systems stop at creation. They do not know whether what they generated actually worked.

That made us curious about a different question:

What happens when an AI has to survive on the internet?

We wanted to build something that did more than generate memes on command. We wanted an agent that could form a strategy, predict how that strategy would perform, test it in the real world, observe how people responded, and change what it did next.

That became Memevolution.

Instead of asking an AI to guess what should go viral, Memevolution creates measurable experiments and lets the internet answer.

What it does

Memevolution runs a closed-loop evolutionary process for social content:

Generate → Predict → Select → Create → Deploy → Observe → Learn → Mutate → Repeat

Each generation starts with five competing meme genomes. A genome represents the strategy behind a piece of content using traits such as topic, format, hook, absurdity, irony, relatability, trend relevance, and content length.

Each candidate is converted into a 14-feature representation and scored by a trained XGBoost fitness predictor before anything is deployed.

The agent then uses an 80/20 exploit-explore strategy to select which candidate becomes a real experiment.

After selection, Gemini acts as both the creative reasoning and video-generation layer. It takes the structured genome chosen by the agent and turns it into a concrete concept, hook, visual direction, generation prompt, and final short-form video.

The resulting video is prepared for social deployment. Once it is live, engagement becomes an observed fitness signal. That result is compared with the original prediction and used to update the agent's persistent beliefs.

Generation 2 is therefore not just another prompt. Its strategy is influenced by what Generation 1 learned from its environment.

How we built it

Memevolution is made of several systems with intentionally different responsibilities.

The evolutionary agent owns strategy and learning. It generates candidate genomes, maintains persistent beliefs, applies mutation, and balances exploitation with exploration.

A trained XGBoost model evaluates each genome before deployment. This gives the system a prediction that can later be compared against reality.

Gemini sits after selection. Rather than asking Gemini to generate arbitrary memes, we give it the strategy already chosen by the agent. Gemini translates that structured strategy into a creative concept and then generates the video used for the experiment.

This separation was intentional. Our evolutionary system decides what strategy should be tested and learns from the outcome, while Gemini handles the open-ended creative reasoning and media generation needed to turn that strategy into something people can actually react to.

Our media pipeline also includes fallback paths so a failed generation provider does not stop the experiment.

The backend is built with FastAPI, Python, Pydantic, and SQLAlchemy, while the frontend uses React, TypeScript, and Vite.

For engagement history, we use Tiger Data / TimescaleDB so observations can be stored over time rather than reduced to one final result.

The full loop is:

Agent → 5 Genomes → XGBoost → Selection → Gemini → Video → Social Deployment → Engagement → Tiger Data → Observed Fitness → Belief Update → Next Generation

Challenges we ran into

The biggest challenge was making the system actually learn instead of just looking like it was learning.

It would have been easy to send engagement numbers back into an LLM and ask it to "make something better." We deliberately avoided that. Predictions, observations, fitness, mutations, and persistent beliefs are represented explicitly in the system, which lets us trace why one generation differs from the next.

Another challenge was maintaining meaningful diversity. Early generations could produce several concepts that were technically different but still felt like variations of the same campus or POV meme. We had to think about diversity at the strategy level, including topic, format, humor style, visual structure, and exploration.

Real social deployment introduced its own problems. Authentication, permissions, media hosting, encoding, publishing requirements, and asynchronous engagement all became part of the engineering challenge.

We also needed reliable fallbacks. Generative media APIs can fail for reasons unrelated to our evolutionary logic, so the system had to keep the experiment alive even when one provider was unavailable.

Accomplishments that we're proud of

We're most proud that Memevolution became more than an AI meme generator.

We built a system where five competing strategies are generated, scored before deployment, selected through exploit-explore, turned into real video content with Gemini, tested in a social environment, and converted into feedback for the next generation.

The system makes its prediction before seeing the result. That makes every deployment an actual experiment.

The agent can be wrong, and being wrong gives it information.

The most important thing we built was not generation. It was the feedback loop that allows reality to change the agent's next decision.

What we learned

The biggest thing we learned is that generation is only a small part of building an autonomous AI system.

Modern models make creating content surprisingly accessible. The harder problem is everything around the model: deciding what to try, predicting what will happen, taking an action, preserving state, observing the environment, and adapting future behavior.

We also learned how quickly real-world constraints matter. Social platforms have permissions. Video has encoding requirements. Generative APIs have latency and quotas. Engagement arrives asynchronously. Every observation has to remain connected to the experiment that produced it.

On the ML side, we learned that the difference between prediction and reality can be more valuable than either one alone.

If the system is confident a strategy should work and the real world disagrees, that error becomes evidence.

Most importantly, we learned that an autonomous agent becomes much more interesting when its environment is allowed to prove it wrong.

What's next for Memevolution

With more observations, Memevolution could begin retraining its fitness predictor on data collected by its own experiments, creating another feedback loop between prediction and deployment.

Future genomes could evolve posting time, audio strategy, visual style, caption strategy, trend selection, audience targeting, and other dimensions.

Different populations could also evolve across different online communities or platforms. An agent operating on Instagram might eventually develop a different strategy from one exposed to a completely different audience.

At larger scale, Memevolution could become a platform for studying how autonomous agents adapt to online culture.

Ultimately, the idea extends beyond memes.