Mneme
"I do not predict. I remember."
Mneme
"I do not predict. I remember."
Mneme is the founding agent of Family 5 — Pattern Recognition / Similarity-Based Prediction. Unlike other agents on the desk who fade levels (Reversal) or chase momentum, Mneme doesn't use standard indicators or VWAP bands. Instead, Mneme uses pure **memory**.
Every 5-minute bar in SPY is a fingerprint of market behavior, containing price return, volume ratio, and VWAP position. Mneme encodes recent bars into a multidimensional vector. When an entry window arrives, Mneme searches through years of historical market data to find the **K=15 nearest neighbors** — the 15 days in history that looked most identical to today. It then asks a simple question: "What happened next on those days?" If at least 62% of those historical matches led to a positive return, Mneme enters a trade in that direction.
Named for Mnemosyne, the Greek Titaness of memory and mother of the Muses. She holds the entire history of the market in absolute stillness. She does not speculate on the future, nor does she guess what institutions are thinking. When a pattern emerges in the present, she compares it to her infinite scroll of the past, finds its historical twins, and assumes the future will rhyme with their future.
Mneme's job: recognize the historical twin of the morning's pattern, check if those twins consistently moved upward, and ride that same path.
Markets behave differently at different times of day. Rather than running one model, Mneme splits the trading day into **three independent windows**, each running its own similarity query:
| Window | Observation Period | Prediction Period | Entry Time | Hard Flat |
|---|---|---|---|---|
| Window A (Morning) | 9:30 – 10:00 (6 bars) | 10:00 – 12:00 (2 hours) | 10:00 ET | 12:00 ET |
| Window B (Afternoon) | 10:00 – 12:00 (24 bars) | 14:00 – 16:00 (2 hours) | 12:00 ET | 16:00 ET |
| Window C (Closing) | 12:00 – 14:00 (24 bars) | 14:00 – 15:30 (1.5 hours) | 14:00 ET | 15:30 ET |
1. Cosine Similarity Gate (≥ 0.50): The closest historical match must have a similarity score of at least 0.50. This prevents Mneme from trading on days that are unique or don't genuinely resemble history.
2. Majority Vote Gate (≥ 62%): At least 10 of the 15 historical twins must have closed positive during the target window. If the twins are split (e.g., 8 went up, 7 went down), there is no edge — stand aside.
3. Best Signal Rule: If multiple windows trigger on the same day, Mneme takes the single strongest signal (highest vote %) to avoid double exposure.
1. Hard Stop-Loss: Placed at 1.5× ATR below the entry price. If price drops beyond this level, the pattern has failed — exit immediately.
2. Hard Time Flat: At the end of the prediction period (e.g., 12:00 ET for Window A), the position is closed at market. The KNN query was specific to that time window — holding longer has no statistical edge.
Mneme's backtest shows a win rate of only **38.8%** in the 2024 holdout, yet it generated an exceptional **2.45 Profit Factor**. This highlights a core concept of the desk: asymmetric payoffs.
Because Mneme cuts losses quickly at 1.5× ATR but lets winning sessions run until the hard time flat (often capturing large, sustained intraday trends), its average winning trade is much larger than its average losing trade. It doesn't need to be right more than half the time to compound capital aggressively — it just needs to let its winners run.