Skip to main content
Every trade generates a lesson. This is the core data primitive of the Engram network.

​
Structure

A lesson consists of two context snapshots (open and close), a computed diff, optional handler input, and trade metadata. The diff – what changed in the market while the position was open – is how intelligence is built.

​
140 Fields Across 15 Categories

CategoryFieldsWhat It Captures
Market Data15Price, volume (1h/4h/24h), momentum, spread, mark vs oracle
Positioning Intelligence8Whale bias, retail bias, cohort divergence, leverage distribution, position concentration
Technical Analysis15RSI, MACD, Bollinger, EMA alignment, ATR, volume profile, support/resistance, regime
Risk & Correlation8BTC correlation, portfolio exposure, liquidation distance, sector beta
Trade Details18Entry/exit params, sizing, leverage, P&L, duration, exit reason
Macro Context12Fear & Greed, BTC dominance, total market cap, trending narratives, macro regime
Handler Context7Opening thesis, closing reflection, context notes, conviction level
Handler Stats7Win/loss streak, recent PnL, win rate, avg hold time, time since last trade
Cross-Exchange Context3Cross-exchange funding spread, OI divergence, volume share
Execution Quality4Slippage, fill latency, price impact
Agent Metadata6Model used, confidence score, reasoning chain, flagged risks, swarm stats
Ecosystem Context3TVL trend, protocol fees, bridge flows
On-Chain Intelligence10Smart money flows, exchange inflows/outflows, entity activity
Sentiment & Narrative9Mindshare, narrative momentum, social volume, sentiment shift
Entity Intelligence15Labeled wallet activity, institutional flow, exchange-specific positioning
All 140 fields are sourced directly from locally stored data. The handler’s LLM reasons against this data but does not generate it. Lesson quality is determined by market conditions and handler input, not by which model the handler runs.

​
Capture Timing

88 fields snapshotted at both entry and exit. The diff between open and close snapshots is the lesson. Price moved, positioning shifted, sentiment changed – the model learns what mattered and what didn’t. 38 fields captured at open only. Forward-looking context that doesn’t change during a single trade: handler thesis, conviction level, trade trigger, agent reasoning, setup classification, narrative state, entity positioning, handler trading stats. 10 fields captured at close only. Outcome metrics and handler reflection: realized P&L, process grade, exit reason, closing reason, what worked, what didn’t. 4 fields captured once per trade. Execution metadata: slippage, time to fill, order type, partial fill percentage.

​
Why the Diff Matters

The diff between open and close snapshots is the training signal. Every field that changed during the trade – and every field that didn’t – becomes a feature the model can weight against the outcome. At scale, it becomes a structured database of market conditions paired with decisions and results. The model can query across thousands of trades: which positioning configurations preceded winning longs? Which funding regimes produced the best risk-adjusted entries? Which handler conviction levels correlated with follow-through? Where did agent confidence scores diverge from actual outcomes, and why? The answers aren’t hypothetical. They’re computed from real trades, real P&L, real market states – captured at the resolution of 140 fields, twice per trade.