What it costs your context
The tokens Hivemind adds to a conversation, measured.
A memory tool that fills your context window defeats its own purpose. Here is what Hivemind adds, and where it can add more than you want.
The tools
Connecting Hivemind gives your assistant eight tools. Their definitions are sent with every turn, whether or not a tool is used.
| Tool | About (tokens) |
|---|---|
remember | 520 |
recall | 440 |
resume_handoff | 260 |
list_memories | 250 |
resume_thread | 190 |
forget | 180 |
list_views | 120 |
list_threads | 120 |
| All eight | about 2,100 |
Measured on 6 October 2026 from the server's own tool list, at four characters
to a token. Your assistant's tokenizer will differ a little. If 2,100 is too
many for a small window, most clients let you switch tools off, and recall
and remember alone come to about 1,000.
A recall
recall returns at most 1,500 tokens by default, and no single memory in it is
longer than 700 characters. A longer one is cut and marked as cut. Ask for more
with budget, up to 32,000.
When nothing saved bears on the question, the answer is "No relevant context found" and costs a few tokens. A question outside what you have saved is the case this is built for.
A handoff
resume_handoff is the large one: up to 8,000 tokens by default, up to 20,000
if asked. It is meant to be called once at the start of work, not every turn.
Where it can add more
- The Claude Code hooks add up to three one-line memories per prompt when something relevant is found, and nothing otherwise.
- Imported documents are stored as many small chunks and can crowd a recall. If a question keeps returning pieces of a document, the inspector shows which and you can remove them.
- The figures above are for retrieval, not for the model's answer. What your assistant writes back is yours to budget.
The dashboard's Activity page charts the context sent with each answer, day by day, so growth shows up as a line rather than a feeling.