[ HIVEMIND FOR RESEARCH ]
A shared-memory substrate for multi-agent experiments. Context compiled under an explicit token budget by a fixed policy, a receipt for every operation, an export for every dataset — with any model, in any harness.
YOUR CONTEXT PIPELINE IS THE UNREPORTED VARIABLE IN EVERY AGENT EXPERIMENT.
“WE USED RAG” IS NOT A METHODS SECTION.
IF YOU CAN'T SAY WHAT THE MODEL SAW,
YOU CAN'T SAY WHY IT ACTED.
Every turn's context is assembled by a fixed, inspectable policy under an explicit token budget — ranked, deduped, trimmed. Set the budget, set the caps per source, and the pipeline stops being a confound.
THE CONTEXT IS A CONTROLLED VARIABLE.
Every compile emits a receipt: what was recalled, what was admitted, what was dropped and why, token counts per source. Exactly what each model saw on each turn — your methods section, generated as you run.
CITE THE CONTEXT, NOT THE VIBES.
Point any number of agents at one workspace and study what shared memory does to collective behaviour. Every write is visible to every agent on its next turn — 150+ turns a minute measured on entry-level storage.
THE HIVEMIND IS THE EXPERIMENT.
This is what provenance looks like — one compile receipt, abridged:
{
"receipt_id": "e9b3c1a0-59f2-11f1-8d2e-0242ac120002",
"operation": "context.compile",
"status": "ok",
"created_at": "2026-07-05T14:32:11+00:00",
"session_id": "session_001",
"turn_id": "14",
"trace_id": "trace_001",
"parent_receipt_id": null,
"inputs_ref": { "session_id": "session_001", "turn_number": 14, "budget_total": 4096,
"recalled_conversation_count": 3, "recalled_memory_count": 7 },
"outputs_ref": { "token_count": 3220,
"item_ids": ["context-briefing-14", "context-status-14",
"context-artifacts-14", "hold-8c2f", "memory-9d41c2"],
"source_counts": { "hivemind": 4, "system": 4, "conversation": 2,
"recalled_conversation": 2, "directive": 1 } }
}Every operation — store, recall, compile, delete — emits one. Queryable by session, turn, or operation via GET /receipts; chained by parent_receipt_id for causal lineage.
Give N agents one mind and vary N. Coordination, division of labour, collective error — every write timestamped and attributed, so you can watch it happen instead of inferring it.
Trace how a fact spreads agent to agent through shared memory: receipts record who stored it, who recalled it, and when. Information flow as data, not diagrams.
Agents that persist across days and sessions — the memory is the state. Checkpoint it with an export, fork it into a fresh project, replay a condition.
Budget, recall depth, and admission policy are request-level variables. Run the same episode under different context regimes and measure what the difference in memory does to behaviour.
When an agent does something surprising, walk the receipt chain backwards: what it saw, where that came from, which earlier operation planted it. Causal lineage, not speculation.
A swarm run's accumulated memory is itself an artifact: GET /export streams the whole store, metadata and timestamps intact, ready for analysis or release.
It slots into a batch harness as one HTTP call per turn: POST /turn carries the agent's input in and the compiled context out, with the receipt inline. Any model, any framework, any language — the service never touches your inference. Between conditions, wipe the workspace or spin up a fresh project; after a run, GET /export streams the entire memory store for your dataset. Storage is a Qdrant cluster you control, so the raw vectors are yours to inspect too.
The API reference documents every endpoint and field in curl, JavaScript, Python, and Go. The main documentation covers the context window model, recall and ranking, and budget mechanics.
14-day trial, card up front, nothing billed today. One afternoon from empty workspace to a swarm with a shared mind.