tradememory-protocol

TradeMemory Protocol is a memory layer for trading agents. Its pitch is that many MCP servers can execute trades but none remember what happened. It records each trade with context such as entry reasoning, market regime, confidence and outcome, then lets the agent recall patterns. The README describes outcome-weighted recall that downweights patterns from old market regimes, and Bayesian confidence plus sample size on each pattern with out-of-sample validation, so suspicious patterns are flagged.

An Evolution Engine goes further: given raw price data, it generates strategy hypotheses with an LLM, backtests them with a vectorised engine (ATR-based stop loss and take profit, long and short, time-based exits), eliminates weak candidates and evolves the rest. The README quotes 22 months of BTC data with a Sharpe ratio of 3.84, 477 trades and 91% positive months; that is the project's own backtest result and should not be read as a promise of live returns. Forex users can sync closed trades from MetaTrader 5.

It exposes 15 MCP tools (core memory, OWM cognitive memory with agent state such as confidence, risk appetite and drawdown, and evolution tools) plus 30 REST endpoints. Install with `pip install tradememory-protocol` and add it to Claude Desktop with `uvx tradememory-protocol`, or use `claude mcp add`, or run it with Docker. The README reports 1,055 passing tests. The licence is MIT, the last push was 2026-10-09, and the README states plainly that it is for education and research, is not financial advice, and that trading carries substantial risk of loss. This review is based on the README and repository data; we have not run it ourselves.