Your trading bot needs therapy, not just tickers

What it does

TradeMemory Protocol is an MCP server that sits between your AI trading agent and its increasingly expensive amnesia. It records every trade decision across five memory layers (episodic, semantic, procedural, affective, and raw trade records), then surfaces relevant past outcomes when your agent faces similar market conditions. It also tracks confidence, drawdowns, and losing streaks — essentially building a therapist’s notes for your bot’s behavioral patterns.

The interesting bit

The “Outcome-Weighted Memory” framework is the unusual angle: recall scoring blends outcome quality, context similarity, recency, confidence, and even “emotional state” (the affective layer). This isn’t just vector RAG with a finance hat — it’s explicitly modeled on Tulving’s cognitive psychology work from 1972, applied to whether your AAPL long was a good idea last earnings season. The SHA-256 chained audit ledger with daily Merkle roots is overkill for most hobby traders, but deliberately designed for MiFID II and EU AI Act compliance submissions.

Key highlights

• 17 MCP tools plus 35+ REST endpoints; runs fully offline with no external network calls
• SHA-256 tamper-proof audit chain: per-record hashes, forward chaining, daily Merkle roots
• Five-factor OWM recall before trading; one remember_trade call writes to all five memory layers after
• Safety rails: pre-trade legitimacy gate (full/reduced/skip), behavioral drift detection, drawdown alerts
• 1,428 tests passing; includes MT5 sync, strategy evolution engine (research phase), and failure taxonomy documenting 11 trading AI failure modes

Caveats

• Empirical validation is ongoing with only n=40 trades so far; the authors note they need n≥100 for statistical significance
• Evolution engine is explicitly in “research phase” with statistical gate pass rate “under optimization”
• TSA timestamping, external anchoring, and zkML proof of inference are listed as not yet implemented in v0.5.2
• “Pro” tier at $29/mo is marked “Coming Soon” with no timeline

Verdict

Worth a look if you’re building production trading agents and need either compliance-grade audit trails or non-trivial memory architecture beyond “stuff everything in a vector DB.” Skip it if you want execution tools — this deliberately doesn’t touch orders, API keys, or wallets. The cognitive-science framing could be genuine insight or academic window-dressing; the n=40 sample size suggests treating the OWM claims as promising but unproven.