The context bill
Most agents remember by re-sending the whole conversation with every model call. That bill grows every day the conversation does, until the history outgrows the context window and the agent forgets anyway. Put in your own numbers.
Re-sending the history
$145,800
$81,000 of it in the last 30 days
With Recall
$7,209
$2,403 of it in the last 30 days
Re-sending costs 20× as much over 90 days.
day 1day 90
- From day 3, each day of re-sending costs more than a day with memory.
- The whole history still fits in a 200,000-token window after 90 days.
How it is worked out
- A day's messages are half the user's and half the agent's, and each user message is one model call.
- Re-sending: each call reads the conversation so far (every earlier day and, on average, half of today), up to the context window, at your model's input price less the cache discount.
- With Recall: each call reads a 2,000-token memory context (a running summary and the facts that matter, from the context endpoint) and the last 10 messages, at the same input price. Recall learns from each message once, at $1.75 per 1M tokens, a message counting as at least 50 tokens. Fetching the context is included in that price.
- Left out of both: output tokens (the same either way), Recall's optional chat questions, and the free credit.
- The calculator runs in your browser and sends nothing. Your numbers are in the address bar, so a link shares them.
Try it on your own agent
New accounts get $5 of credit and a sample conversation to try, no code needed. Adding Recall to an agent takes two calls: store each message, and ask for the context before each model call.