Resident, not replayed
The run is held once and referenced by handle. Turn forty does not pay to re-read turns one through thirty-nine, which is where a long agent loop quietly spends its budget.
O(1) per turnWhat sits in the aperture
A context window is not a prompt you keep re-sending.
Every turn, every tool result and every document stays where the model left it. Recall addresses them directly, so a long run costs what a short one does.
Addressable, not a blob
Ask for the third tool call or the paragraph a claim came from and get exactly that back. The window is an index, not one long string you slice by hand.
Span-level readsEvictable on your terms
You decide what ages out and what is pinned for the life of the run. Nothing disappears because a buffer filled up while you were not looking.
Explicit policy1Mtokens addressable per run
6 msmedian span fetch
0re-reads on turn forty
94%fewer prompt tokens billed
Invented product, invented figures. Nothing here was measured.
Three calls
Open it, write to it, point the model at it.
Start a window
One call per run returns a handle. Everything after that is written against it, including the turns your framework would otherwise be concatenating for you.
Append as you go
Turns, tool results and documents go in as spans with whatever metadata you want to query them by later. Writes are append-only, so a replay is exact.
Hand over the handle
The model reads through the window instead of through your prompt. You send the question and the handle, and nothing else travels.
Eviction policy
You decide what the window keeps.
Nothing ages out on its own schedule. Set the policy once per run and the window holds to it.
Retention
Applies to the whole run
Pin tool resultsNever evict a tool call or its output
Pin cited spansKeep anything an answer has pointed at
Summarise on evictLeave a stub behind instead of a hole
The run
24 spans · 1,100K written62%in use
- Pinned0K
- Kept0K
- Stubs0K
- Evicted0K
OldestTool resultCitedNewest
Plans
Priced per window, not per token you resend.
$0one window at a time
- 128K addressable per run
- Span reads and pins
- Seven day retention
$40per seat, per month
- 1M addressable per run
- Policy per run, not per org
- Replay from any append
- Ninety day retention
Talkannual, on your metal
- Unlimited windows
- Your storage, your keys
- Air-gapped deploys
Questions
The ones worth asking.
Is this a vector store?
No. A vector store is for finding things you did not know you had. A window is for holding things you already put there, in order, and reading them back by address. You would sensibly use both.
What happens when a run exceeds the budget?
The policy you set decides. Pinned spans stay, the oldest unpinned spans go, and with summarise-on-evict a stub is left in their place so the order of the run is still readable.
Does the model need to know about Recall?
Only enough to be handed a window id. Everything else is the runtime's job, so the same window works across whichever model you route to that day.