io.github.mnemox-ai/tradememory-protocol

Getting Started | Use Cases | API Reference | OWM Framework | Limitations | 中文版

TradeMemory remembers what it cost. It is an open-source, local-first memory and brake for AI trading agents: it pulls in your fills, finds where your own history loses money, puts those losing trades in front of your agent before the next order, and can refuse an order that breaks rules you set, before it reaches the broker.

Brokers now let AI agents trade over MCP, with different guardrails: Webull and tastytrade set size or buying-power limits, Interactive Brokers only lets the agent draft an order for you to submit, and Robinhood and Public set no cap you can impose on an external agent (StockBrokers, 2026-09-30). The caps are fixed numbers. None of them looks at how your own past trades went.

Start with your own history

pip install tradememory-protocol

# Hyperliquid: public fills, no key needed
tradememory sync hyperliquid --address 0xYourAddress

# Alpaca: read-only calls with your own keys, kept in a local file
tradememory sync alpaca --env-file ~/.secrets/alpaca.env

Each closed trade is stored in memory once; running it again stores only new trades. Then it prints where your history loses money. The format, with illustrative numbers:

After 2 losses in a row (20 trades):
  5 of them (25%) were 1.5x your usual size or more.
  All 20 won 60% and made -$1,500.
  The 5 sized-up trades won 20% and made -$1,700.

Median hold: winners 1.5h, losers 9.0h.

These are descriptive statistics of your own past trades, not advice about the next one. Hyperliquid's API serves only an address's recent fills, not its whole history, so the sooner it is synced, the more history is kept. Other venues: MT5 and Binance spot sync scripts are in scripts/, and any agent can record a trade with remember_trade.

Before the next order

recall_memories(order="losses_first") puts every losing trade ahead of the rest, the bigger losses in similar conditions first, with their size and P&L. Besides the symbol's most recent trades, it searches the symbol's recent losing trades, so an older loss is not crowded out. The server tells connected agents to call it before proposing a trade. The default order ranks better outcomes higher (by R multiple, where one was recorded) and keeps at least 20% losses in the list.

Connect your agent

pip install tradememory-protocol

Add to Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "tradememory": {
      "command": "uvx",
      "args": ["tradememory-protocol"]
    }
  }
}

Then tell Claude: "Record my AAPL long at $195 — earnings beat, institutional buying, high confidence."

# Claude Code
claude mcp add tradememory -- uvx tradememory-protocol

# From source
git clone https://github.com/mnemox-ai/tradememory-protocol.git
cd tradememory-protocol && pip install -e . && python -m tradememory

# Docker
docker compose up -d

Full walkthrough: Getting Started (Trader Track + Developer Track)

Put a brake in front of your broker (preview)

The proxy extra runs TradeMemory between your agent and your broker's MCP server (Alpaca today). Every tool is forwarded unchanged except the order-placing ones, which Mnemox Control evaluates against a policy you own before they reach the broker. Every evaluation, allowed or refused, is recorded and chained into the audit log. An allowed order comes back with your losing trades from similar conditions first, and tradememory sync alpaca later fills in how each forwarded trade ended, in R when the entry carried a stop. The agent you already have keeps working; only one line of its MCP config changes.

pip install "tradememory-protocol[proxy]"   # Python 3.12+
tradememory proxy init --account-id <your Alpaca account id> --symbols AAPL,MSFT
tradememory proxy doctor --env-file ~/.secrets/alpaca-paper.env   # checks the live tool names and the account id
tradememory proxy config                                           # prints the MCP client entry that replaces the direct Alpaca one

Refused by default: symbols outside your list, orders above your notional and position limits, entries without a bracket stop, any new order after your daily-loss or drawdown limit, cancelling the protective stop of an open position, any tool the brake has not classified, and everything while you have run tradememory proxy halt FULL_HALT. Never blocked: closing a position. Orders at or above approval_notional wait for tradememory proxy approve <intent_id> --terms <fingerprint>, which approves exactly the terms you read; the agent retries with the same client_order_id and the same terms, and the proxy forwards it at most once. evaluate_order returns the same decision without placing anything, for pre-checks and for advisory layers in other frameworks. Anything the brake cannot evaluate (a dead quote feed, an unknown asset, an order type the policy does not cover) is refused, not passed through.

Status: tested end-to-end against a stateful fake of Alpaca's MCP server (tests/proxy/), and run against a real Alpaca paper account on 2026-10-01: three refusals (symbol not on the list, entry without a stop, notional over the limit), one allowed one-share bracket order that reached the broker, and one retry with the same client_order_id that was answered from the record without a second order. Options, order replacement, stop-limit and trailing orders are refused rather than evaluated. Your broker keys go only to the broker process the proxy starts; TradeMemory never stores them.

Walkthrough with the real outputs: docs/recipes/alpaca-brake.md.

Three ways it is used

How it works

1. Recall — Before trading, retrieve past trades weighted by outcome quality, context similarity, recency, confidence, and emotional state (OWM Framework)
2. Record — After trading, one call to remember_trade writes to five memory layers: episodic, semantic, procedural, affective, and trade records
3. Reflect — Daily/weekly/monthly reviews detect behavioral drift, strategy decay, and trading mistakes
4. Audit — Every decision is SHA-256 hashed at creation and chained to the one before. Export anytime for review

MCP Tools

REST API: 35+ endpoints for trade recording, reflections, risk, MT5 sync, OWM, evolution, and audit. Full reference →

Get help connecting

Running an agent against a broker, or want your history synced and read? Open a brake-integration issue or book 30 minutes. Setup help is free, and the people who use it decide what gets built next.

Audit trail

Every trading decision your agent makes — including decisions not to trade — is recorded as a Trading Decision Record (TDR). Per-record SHA-256 content hashes are linked into a forward-chained audit ledger; every UTC day is summarised by a Merkle root which itself chains across days. Tampering with any historical record invalidates every subsequent link.

Rules that require decision records bind investment firms, not retail users. Under the EU AI Act, the Annex III high-risk logging obligations were postponed to 2 December 2027, and ESMA's February 2026 supervisory briefing on algorithmic trading states that AI-based algorithmic trading is currently excluded from the high-risk scope. The table shows which TradeMemory features map to those texts if and when they apply to you. It is not a compliance claim.

# Verify a single record hasn't been tampered with
verify_audit_hash(trade_id="MT5-7047640363")
# → {"verified": true, "chain_entry": {"sequence_num": 42, ...}}

# Walk the entire chain (or a slice) end-to-end
verify_audit_chain(from_seq=1, to_seq=None)
# → {"verified": true, "checked_count": 1284, "first_break_at": null}

# Daily Merkle root — single 32-byte anchor over every TDR for that day
get_daily_root(date="2026-05-14")
# → {"verified": true, "root_hash": "a05544...", "record_count": 18}

# Bulk export
GET /audit/export?strategy=VolBreakout&start=2026-03-01&format=jsonl

Daily roots are timestamped by an RFC 3161 authority by default since 0.5.3. Not built: signing records with a private key, anchoring to a public log, proving that nothing was left out. See LIMITATIONS.md.

Security

• Memory server (default). Never places orders and never asks for broker keys. Records and recalls only, in a local SQLite file.
• Sync. tradememory sync hyperliquid reads public data with no key. tradememory sync alpaca makes read-only calls with keys from a file on your machine. Synced trades stay in your local database.
• Brake (proxy extra). Forwards the orders your policy allows to your broker's MCP server. Your broker keys are passed only to the broker process the proxy starts; TradeMemory never stores them.
• Outbound calls. RFC 3161 timestamping of daily audit roots, a 32-byte hash with no trade data (on by default; TRADEMEMORY_TSA=off turns it off). tradememory sync calls the venue you name. The evolution tools read public Binance market data (api.binance.com) when an agent calls them, and evolution calls the Anthropic API only if you set ANTHROPIC_API_KEY. Replay calls DeepSeek by default (or Anthropic), and only with that provider's key set. If you install sentence-transformers for hybrid recall, it downloads its model from Hugging Face on first use. The MT5 and Binance sync scripts in scripts/ read their credentials from your environment, and the MT5 one posts trade summaries (symbol, prices, P&L) to a Discord webhook only if you set DISCORD_WEBHOOK_URL.
• Tamper-evident, not tamper-proof. Every record is hashed and linked to the one before, with daily Merkle roots. Changing a record breaks the chain; someone who can rewrite the whole database can rebuild it.

Research Status

TradeMemory's OWM framework is grounded in cognitive science (Tulving 1972) and reinforcement learning (Schaul et al. 2015). Current status:

• OWM five-factor scoring: implemented and tested (see the CI badge)
• Statistical validation: DSR, MBL implemented (Bailey-de Prado 2014)
• Audit trail: SHA-256 tamper-evident TDR
• Evolution engine: research phase (strategy generation works, statistical gate pass rate under optimization)
• Hybrid recall: OWM-only mode active, vector fusion available when embeddings configured
• Empirical validation: ongoing (n=14 trades, target n>=100 for statistical significance; at n=14 the confidence intervals are too wide to conclude anything - see validation/final_verdict.md)

Documentation

Contributing

See Contributing Guide · Security Policy

MIT, see LICENSE. The optional proxy extra installs Mnemox Control, whose engine is AGPL-3.0-only (a commercial license is available); TradeMemory itself stays MIT. For educational and research purposes only. Not financial advice.