An internal R&D project that combined MemRL episodic memory, Kimi K2.5 reasoning, and autonomous web scraping to build a self-evolving esports prediction agent achieving a 66% win rate and 12% ROI.
Sports betting markets are notoriously efficient. Bookmaker odds are set by sophisticated models with access to vast historical datasets, real-time performance telemetry, and market signals from millions of participants. Consistently beating the closing line — the final odds before an event starts — is one of the hardest problems in predictive analytics.
Esports compounds this difficulty. Unlike traditional sports, esports titles evolve constantly — game patches shift the meta, roster changes happen frequently, and map or character pools introduce combinatorial complexity that renders static models obsolete within weeks. Historical performance data is fragmented across multiple platforms, inconsistently formatted, and often incomplete. The volatility of the esports landscape means any model built on fixed assumptions decays rapidly.
Corporate Agents set out to build an autonomous betting agent as an internal R&D project — not to operate a gambling business, but to stress-test our agent architecture against a domain where the feedback signal is immediate, unambiguous, and financially quantifiable. If an agent can learn to make profitable decisions in a market designed to extract value from participants, the same architecture can be applied to any domain that requires continuous adaptive decision-making under uncertainty.
The agent was equipped with web scraping tools to autonomously collect match data, odds movements, team rosters, patch notes, and historical results across multiple esports titles. Rather than relying on a static dataset, the scraping pipeline ran continuously — ingesting new data as matches were played and odds shifted. This gave the agent access to a live, evolving picture of the competitive landscape rather than a frozen historical snapshot.
Before placing any forward-looking predictions, the agent ran extensive backtesting across thousands of historical matches. The backtesting engine evaluated candidate algorithms against known outcomes — measuring not just win/loss accuracy but edge over closing odds, variance, drawdown, and return on investment. This phase identified which statistical patterns and feature combinations carried genuine predictive signal versus those that were artefacts of overfitting or survivorship bias.
The core innovation was the integration of a MemRL (Memory-Enhanced Reinforcement Learning) pipeline for continuous, non-parametric learning. MemRL is a framework that enables AI agents to self-evolve at runtime by applying reinforcement learning to an external episodic memory — without modifying the underlying model's parameters.
The agent's memory was structured as triplets: (Intent Embedding, Experience, Q-Value). Each triplet captured the semantic context of a betting decision, the strategy applied, and the learned utility of that strategy based on actual outcomes. When evaluating a new match, the agent used MemRL's two-phase retrieval mechanism:
After each match resolved, the agent received an unambiguous reward signal (win/loss, actual return vs. expected return) and updated the Q-values of retrieved memories using an exponential moving average:
Q_new ← Q_old + α(reward - Q_old)
This mechanism meant the agent continuously learned which patterns were genuinely predictive and which were noise — even when two experiences appeared semantically identical, the Q-value distinguished strategies that worked from those that merely looked plausible. Unlike traditional fine-tuning, this approach carried zero risk of catastrophic forgetting — the LLM backbone remained frozen while the memory layer evolved.
The reasoning layer was powered by Kimi K2.5, selected for its strong performance on complex multi-variable reasoning tasks. The model received curated context from the MemRL pipeline — relevant historical experiences weighted by proven utility — and synthesised a final betting decision. The frozen model architecture meant decision quality improved purely through better memory curation, not through expensive and unstable weight updates.
The agent operated in a closed feedback loop: scrape → analyse → predict → observe outcome → update memory weights → repeat. Each prediction cycle refined the MemRL memory, gradually shifting weight away from patterns that failed to deliver edge and toward those with demonstrated predictive value. Over successive iterations, the agent's effective strategy set was pruned and sharpened by real-world outcomes rather than training data assumptions.
| Metric | Value |
|---|---|
| Win Rate | 66% on esports matches |
| ROI | 12% return on investment |
| Learning Method | Non-parametric (no model fine-tuning required) |
| Prediction Domain | Esports (multiple titles) |
The 66% win rate exceeded the break-even threshold required to generate positive ROI against bookmaker margins, and the 12% ROI demonstrated sustained edge — not a short-term anomaly. The MemRL pipeline's ability to dynamically re-weight pattern importance based on match outcomes was the critical differentiator. Static models trained once on historical data degraded as the meta shifted; the agent's memory-based approach adapted continuously.
| Component | Technology |
|---|---|
| Decision Engine | Kimi K2.5 (frozen parameters) |
| Learning Pipeline | MemRL — episodic memory with Q-value reinforcement |
| Data Collection | Autonomous web scraping agents |
| Backtesting | Custom engine across historical match datasets |
| Memory Structure | Intent Embedding + Experience + Q-Value triplets |
| Retrieval | Two-phase: semantic similarity → value-aware selection |
| Feedback Signal | Match outcomes (win/loss, actual vs. expected return) |
| Update Rule | Exponential moving average Q-value updates |
Disclaimer: Past performance in predictive analytics does not guarantee future results. This case study documents an internal R&D project and does not constitute financial advice or endorsement of gambling. All metrics reported reflect observed performance during the testing period.