mnemox-ai Trading Memory
Overview
TradeMemory implements a cognitive memory architecture for trading agents. Every trade is stored with full context (market conditions, strategy, reasoning, confidence) and recalled using Outcome-Weighted Memory (OWM) — a scoring system that surfaces winning trades in similar contexts first.
This is not a trade journal. It's a memory system that learns which past experiences are most relevant to current decisions.
Architecture: 3-Layer Pipeline
L1: Raw Trades → L2: Pattern Discovery → L3: Strategy Adjustments
- L1 (Episodic): Every trade stored as-is with full context. The ground truth.
- L2 (Patterns): Behavioral patterns discovered from L1 data. Disposition effect, session biases, strategy correlations.
- L3 (Adjustments): Concrete strategy adjustments derived from L2 patterns. Parameter changes, rule modifications, strategy retirement.
Outcome-Weighted Memory (OWM) — 5 Memory Types
1. Episodic Memory
Raw trade events. Each record contains: symbol, direction, entry/exit, P&L, strategy, market context, reflection, timestamp.
When to write: After every completed trade. When to read: When recalling past trades for decision-making.
2. Semantic Memory
Strategy knowledge base. Aggregated understanding of what works: "VolBreakout performs best in London session with ATR > $40" is semantic memory.
When to write: Automatically updated when trades are stored via remember_trade.
When to read: When evaluating whether a strategy fits current conditions.
3. Procedural Memory
Behavioral baselines. Tracks execution patterns: average hold times per strategy, lot sizing consistency, stop loss adherence, entry timing precision.
When to write: Automatically computed from trade history. When to read: During behavioral analysis and daily reviews.
4. Affective Memory
Emotional/confidence state. Tracks: current confidence level (0-1), drawdown percentage, win/loss streaks, risk appetite, tilt indicators.
When to write: Updated after every trade and during daily reviews. When to read: Before entering trades (am I on tilt?), during risk checks.
5. Prospective Memory
Active trading plans. Future-oriented: "If XAUUSD breaks above 5200 with ATR confirmation, go long." Plans have entry conditions, exit conditions, risk parameters, and expiry dates.
When to write: When creating trading plans. When to read: When checking if current market conditions match any active plans.
OWM Recall Scoring
When you query recall_memories, results are scored by:
| Factor | Weight | Description |
|---|---|---|
| P&L Outcome | 40% | Profitable trades score higher. Magnitude matters. |
| Context Similarity | 30% | How closely the recalled context matches the query context |
| Recency | 20% | Recent trades weighted more (exponential decay) |
| Confidence Calibration | 10% | Trades where confidence matched outcome score higher |
Why outcome-weighted? Traditional trade journals treat all trades equally. OWM amplifies signal from successful decisions in similar contexts. If you've profited 5 times trading London session breakouts, those memories surface strongly when you're evaluating the next London session breakout.
MCP Tools Reference
Core Memory (2 tools)
| Tool | Use Case |
|---|---|
get_strategy_performance |
Aggregate stats: win rate, PF, P&L per strategy |
get_trade_reflection |
Deep-dive into a specific trade's reasoning |
OWM Cognitive Memory (6 tools)
| Tool | Use Case |
|---|---|
remember_trade |
Full OWM store: writes to all 5 memory layers |
recall_memories |
OWM recall: scored by outcome, similarity, recency, calibration |
get_behavioral_analysis |
Procedural memory: disposition ratio, hold times, Kelly criterion |
get_agent_state |
Affective state: confidence, drawdown, streaks, risk appetite |
create_trading_plan |
Prospective memory: entry/exit conditions, risk parameters |
check_active_plans |
Evaluate active plans against current market conditions |
Best Practices
When to Record
- Always record after a trade closes, not while it's open
- Include the full market context — session, volatility, trend state
- Write an honest reflection — why you entered, what you expected, what happened
- Set confidence before seeing the result (not after)
When to Recall
- Before entering a trade: "Have I been in this situation before? What happened?"
- During daily review: "What patterns emerge from this week's trades?"
- After a loss: "Have I seen this failure mode before?"
When NOT to Recall
- Don't recall mid-trade to justify holding a loser
- Don't recall to confirm a decision you've already made (confirmation bias)
- Don't over-query — if you're recalling 20 times a day, you're procrastinating, not trading
Common Mistakes
| Mistake | Why It's Bad | Fix |
|---|---|---|
| Recording without context | Useless for recall — can't match future situations | Always include session, volatility, trend state |
| Setting confidence after seeing P&L | Destroys calibration scoring | Set confidence at entry, before outcome is known |
| Ignoring affective state | Trading on tilt leads to revenge trades | Check get_agent_state before every session |
| Never running daily reviews | Behavioral drift goes undetected | Run /daily-review at end of each trading day |
| Storing paper trades as real trades | Pollutes performance metrics | Tag paper trades separately or use a different database |
Data Flow
Trade Closes
↓
remember_trade() → Episodic (raw event)
→ Semantic (strategy knowledge update)
→ Procedural (behavioral baseline update)
→ Affective (confidence/streak update)
→ Prospective (check active plans)
↓
recall_memories() ← OWM scoring
↓
Next Trading Decision