TradeMemory Protocol

TradeMemory stores what an agent did and why. Five memory types are recalled with weighting by realized outcome; a five-factor legitimacy gate plus drawdown and losing-streak rails run before a position is opened; and every Trading Decision Record is content-hashed and forward-chained as chained_hash = SHA256(prev_chained_hash || content_hash), with each UTC day summarised by a Merkle root that chains across days — checkable through the verify_audit_hash, verify_audit_chain and get_daily_root tools. Storage is a local single-tenant SQLite file and nothing leaves the machine except an optional RFC 3161 timestamp call, disabled with TRADEMEMORY_TSA=off. The server is registered in the official MCP registry, and the repository also ships packaged agent skills under .skills/. Self-reported test counts across the README, site and skill files disagree with each other; a fresh clone of master on 2026-08-17 contains 1,452 test functions across 81 files.

Its most unusual property is a LIMITATIONS.md that publishes the project’s own failed validation rather than hiding it: “Phase 5 rigorous validation: INVALID. 100 experiments (2 symbols x 1h x 50 grid strategies x 5 agents) showed that the CalibratedAgent skipped 97% of trades, so the apparent drawdown reduction came from not trading, not from skill. 0/100 DSR PASS.” The same file marks the BOCPD detector and the DQS score as dead and puts the empirical basis at n=40 trades against a target of n≥100. Performance figures quoted elsewhere in the repository are self-reported simulation results over strategy and agent grids, not live trading, and the README’s claim of production use by traders and EA systems is self-reported and unverified.

The practical caveats are documented in the same place: no authentication, API keys, RBAC or rate limiting; MetaTrader 5 as the only broker connector; and a storage layer mid-migration, with SQLite as the production write path and a parallel PostgreSQL/Alembic setup not yet in production. One contributor accounts for 366 of the 367 commits, and as of August 2026 the maintainer declares the project “Feature-complete, in maintenance mode - bug and security reports are still reviewed; no new features or hosted service are planned.” Stars — 1,410 in under six months for an organisation created in February 2026 — run well ahead of installs at roughly 890 PyPI downloads a month, so weigh the latter. The maintainer separately sells a paid analysis of a customer’s own trade history, scoped to descriptive statistics with no signals or advice.