Context served
The console's usage page shows what your agents were given, beside the size of what you store.
The same figures are at GET /me/context-usage for any member or agent with the read scope.
The figures compare what your agents were given with the size of what is stored here; they do not measure what your agents loaded before, or money.
The page shows one period at a time: choose Today, 7 days or 30 days (calendar days in UTC; 7 days is shown first). It starts with one line saying about how many tokens your agents received in that period, then a table with a row for everyone together, people's own requests through the console included, and a row for each agent. Select a column's header to sort the agents by it. An agent with no requests and no active days in the period is left out. Counts are shortened (48K, 1.5M); hold the pointer over one to see the exact figure. How the figures are counted is under How we count on the page.
For context assembly (assemble_context, POST /context) and memory recall (recall_memory,
POST /memory/recall), the table shows:
- Requests: the requests answered, including an answer that found nothing relevant.
- Refused: the requests refused because the month's retrieval calls were spent. No other refusal is counted here.
- Served total: the tokens returned in those answers. For context it is the answer's
used; for recall, each returned entry's statement plus 16 tokens for the entry. - Median: the typical answer size, approximate. Sizes are counted in buckets with upper bounds of 250, 500, 1,000, 2,000, 4,000, 8,000, 16,000 and 32,000 tokens; the median is the middle of the bucket that holds the middle request. When that request is larger than 32,000 tokens, the page says "over 32K". When the answers held nothing (0 tokens served), no typical size is shown. The refused column appears only when a request was refused in the period.
- Stored: the tokens in your documents' current revisions and your current memory entries, at the time of reading. Earlier revisions and superseded entries are not counted. The figure is read at most once a minute and is the same for every period.
These requests are fewer than the plan's Retrieval calls under Quotas. A retrieval call is counted for every search, context and memory recall request when it starts, so a search, or a request that finds nothing or fails, counts there but not here. Under that quota the Usage page shows this calendar month's retrieval calls split into searches, context requests and memory recalls, with how many context and recall requests were counted but not answered.
A person's requests through the console are shown as "People, through the console". The page shows how many times larger the stored material is than the average answer only when the stored figure is the larger one.
Imported instruction files
If you brought your repository's instruction files into Retiform with the Claude Code plugin's
import, the page shows one more figure: how many of those files are stored and the tokens in their
current revisions. It is the nearest thing Retiform knows to what a session loaded from those files
before. In the report it is imported, with documents and tokens.
- What is counted: the documents the import uploaded. The import names each file after the project your repository is linked to, and that name is how Retiform tells them apart, so a file counts whether or not a decision was recorded from it. They are part of the stored figure and are read with it, at most once a minute.
- What changes it: uploading a file again counts its new revision, and a deleted file is no longer counted. Files stop being counted if the project they were imported for is deleted.
- What it is shown beside: the typical recall answer in the chosen period. That figure is your whole team's, not one project's, and a session may recall more than once. The page says how many times larger the files are only when they are the larger, and not when the typical answer is over 32K.
A team that has imported nothing has no imported in its report, and the page says in one line that
importing the instruction files makes the figure appear.
How often agents use it
For 7 and 30 days, the same table also shows, for each agent and for the whole team, on how many days each agent:
- was active: made any request to Retiform, even one that only connected. An agent that is connected and never reads or writes is listed with its days and zeroes;
- read: had an answered request for context or for memory recall. A refused request is not a read, and plain search is not counted;
- wrote: recorded or superseded a memory entry. Deleting, confirming or reviewing an entry is not a write;
- did neither on a day it was active;
- read first (or no write): its first read that day came before its first write, or it read and did not write. A day on which it wrote first counts as a day with a read and a write, not as one on which it read first.
The whole team's figures add up each agent's days, so they count agent-days, not team days. Where a client declares a session label on its writes, the page also shows, for each label, the days with a write and the number of writes. The label is the client's own text and Retiform does not check it.
The page sets no target and does not score, rank or compare agents, people or teams. An agent that is connected but never reads has nothing telling it when to, so a row of active days with no reads is a reason to check that agent's instructions or setup.
The unit. Every size is in tokens as counted by kg-utf8-quarter-v1: a quarter of the text's
UTF-8 bytes, rounded up. It is Retiform's own estimate, not any model's tokenizer, and it over-counts
English text, so a model's own count is usually lower.
What is kept. For each agent and day, only the counts and sizes above, the time of its first read, and a mark that it was active; never the questions, answers or stored text. Writes and their labels are not copied: they are counted from your team's own audit trail each time the page is read. The records are kept for 35 days and are removed when your team's data is erased.