Usage

What Memory Soda itself spent on the currently selected project: every model call and embedding batch it made, what they cost, how long they took, and how memory grew over the window.

This is the platform’s own spend. The tokens your application reports on messages (tokens on addMessage) are shown separately as stored message tokens and never counted as cost.

Where the numbers come from

Every unit of work that costs money or time writes one row to the usage_logs table:

kind	One row per	Examples of stage
llm	text or structured completion	extract_episode, extract_graph, resolve_contradictions, summarize, synthesize, reply
embed	embedding batch (≤100 texts)	embed_summary, embed_entities, embed_facts, embed_query
span	timed operation, no model	recall, episode, semantic, compact, http

Each row also records where it came from: source (api for an SDK key, dashboard for a signed-in session, worker for a background job), the API key, the user, the dataset, thread and episode, and a request_id shared by every row one request or job produced. Anything else (message counts, fact counts, HTTP status) lives in a meta JSON column.

Cost is never stored. It is computed when you look, from the row’s service + model and a price table in apps/api/src/lib/usage.ts (USD per million tokens). A model with no entry shows as unpriced rather than $0. Gemini’s embedding API returns no token count, so embedding cost is estimated from characters (÷ 4).

Logging is best-effort and off the request path: rows are buffered in memory and written in one batch every five seconds. A crash can lose a few seconds of rows; a request is never slowed.

The page

Filters (top row): window (7 / 30 / 90 days, 12 months, which also sets the bucket to day / week / month), dataset, source, stage, and service / model. Options are taken from the current result, so a filter only offers values that exist.

Overview

Stat tiles for the window: cost, tokens (in / out), calls and error rate, p95 / p50 latency, live and invalidated facts, episodes by status, threads, and stored message tokens. Below them, per-bucket bars for cost, tokens, calls and errors, and by dataset / by API key tables answering “who is costing me”.

Breakdown

One table, grouped by your choice of stage, operation (the route or worker job), service / model, source, or kind, with calls, errors, tokens, cost and p50 / p95 latency. Spans are timing-only and show no cost.

Memory

New threads, messages, episodes, facts and entities per bucket.

Logs

The raw rows, newest first, with the same filters. Expand a row for its request id, dataset, thread, episode, error and meta.

API

Dashboard-only, session-authenticated; there is no SDK counterpart.

GET /dashboard/projects/:projectId/usage?from=…&to=…&bucket=day|week|month
    [&dataset=&source=&operation=&stage=&kind=&service=&model=&apiKeyId=]
GET /dashboard/projects/:projectId/usage/logs?limit=50[&cursor=<createdAt>]  (+ same filters)

from / to default to the last 30 days.

Adding a model or provider

The log is provider-neutral: a new client logs with its own service and model, and the page groups by them automatically. To price it, add one row to PRICES in apps/api/src/lib/usage.ts. Nothing else changes.