TraderMemosAI features
Optional screenshot fill extraction and an LLM trade coach — any OpenAI-compatible API, your keys, stored on your server.
Both AI features are off by default and entirely optional — TraderMemos works fully
without them. They talk to any OpenAI-compatible chat/completions endpoint: OpenAI,
OpenRouter, Azure OpenAI, or a local server like Ollama / LM Studio. You bring the keys.
Configure both in Settings → AI. Two things to know up front:
- Settings are server-wide (one config per instance, not per user).
- API keys are stored on the server only — after saving, the browser sees just a masked hint.
Screenshot scan (fill extraction)
Extracts fills from a broker-app screenshot so you don't retype them. In the New Trade drawer, Scan to fill takes one or more PNG / JPEG / WebP screenshots — or a CSV / JSON fill file — and stages them in a review overlay before anything touches the form:
- each file gets its own row and progress while the model reads it;
- when all have settled, per-symbol cards show the fills, average entry / exit and FIFO net P&L (multi-symbol screenshots work);
- the form only changes when you press Fill form. A scan that read nothing offers Close instead, with the model's notes as the explanation.
The server does the careful parts. Fill times are read in the timezone the screenshot shows
(a label like 美東 beside the time), else your market timezone, and a missing year is
filled in without putting a fill in the future. Cancelled, rejected and unfilled orders are
skipped, and long / short comes from the fills themselves. When the provider is overloaded,
the server retries twice and then answers 503 ocr_busy with a "try again" message.
Setup in Settings → AI → Screenshot scan:
| Field | Notes |
|---|---|
| Enabled | Turn the feature on |
| API base URL | e.g. https://api.openai.com/v1, https://openrouter.ai/api/v1, http://localhost:11434/v1 |
| Model | Must support vision — default gpt-4o-mini. Fetch models lists what your endpoint offers |
| API key | Leave blank on later edits to keep the current key |
| Custom prompt | Optional — blank uses the built-in fill-extraction prompt |
Use Test to verify connectivity before saving. Images are capped at 10 MiB
(TM_OCR_MAX_BYTES) and requests time out after 90 s (TM_OCR_VISION_TIMEOUT_SEC) — slow
local models may need a higher timeout, and your
reverse proxy must allow ≥ 120 s.
Trade coach
On any trade's detail view, Ask AI sends the trade to your model — fills, P&L, R-multiple, MAE/MFE, setup, emotion, tags, your notes — plus your state as of the entry: that day's and week's prior P&L, win/loss streak, drawdown, and the emotions on your last few closes. The review comes back as:
- coaching notes, streamed in one at a time as the model writes them;
- one next action — a single checkable instruction for the next trade;
- a per-trade execution score (entry, exit, risk, tempo on the same 0–100 scale as Reports).
Reviews are saved: reopening the trade shows the last one (Saved review from …), and the button becomes Regenerate. When the coach is off or unreachable, the panel falls back to built-in rule-based notes.
Setup in Settings → AI → Trade coach: same fields as above (any chat model works — vision not required), plus an optional System prompt to change the coaching style.
On mobile, the model row opens a searchable picker sheet grouped by vendor. For screenshot scan it hides models that can't read images (behind Show all models), and you can type any model id the endpoint doesn't advertise.
Headless defaults via environment
For scripted deploys you can pre-seed both features with TM_OCR_* / TM_COACH_* variables —
see Configuration. Env values act as defaults only:
once settings are saved in the UI, the database wins.
Privacy
Trade data and screenshots go only to the endpoint you configure. Point it at a local model and nothing leaves your machine.