N agents, N memories
Each agent keeps its own partial picture in its own vendor silo. The same lesson gets re-learned five times, and the sixth agent still makes the mistake.
Fleets · shared memory · one brain
Many agents. One brain.
A fleet is only as smart as what its agents share. CueCrux gives every agent, on every model, on every machine, the same receipted memory: write once, recall anywhere, govern everything.
The fleet memory wall
Why fleets plateau
Context stuffing does not scale
The obvious fix, paste everything into every prompt, measurably backfires: at production scale more context makes strong models worse and multiplies the bill by the fleet size.
Nobody owns the truth
When memory lives inside each vendor, "what does the fleet know?" has no answer, no export, and no audit trail. The org forgets what its own agents learned.
The shape of the fix
One spine under the whole fleet
Every agent, whatever model runs it, talks to the same daemon: facts, sessions, handoffs, decisions and receipts in one place. A lesson learned by one agent at 09:14 is what every agent knows at 09:15, and each write is signed, so the shared brain is also an audit trail.
- claude-fleet-02
- claude-fleet-07
- gpt-ops
- codex-ci
- operator/myles
one shared memory
facts · sessions · handoffs · decisions · receipts
- laptop
- CI runner
- prod server
Write once, recall anywhere: what one agent learns at 09:14 is what every agent knows at 09:15. Sync across machines rides passports (Pro); each hop carries a receipt.
Memory sync
How the brain stays whole
Facts, not transcripts
Agents store distilled facts (about 12 tokens each) instead of replaying whole conversations. Recall is a query with a hard token budget, not a paste.
Handoffs between agents
One agent bundles its working state (facts, open questions, work items) and the next agent accepts it. Cross-model, cross-session, receipted at both ends.
Portable by design
The whole memory exports to a signed bundle and verifies offline. Moving machines, or leaving entirely, is a command rather than a negotiation.
Synced across machines
Laptop, CI and servers share the brain through hosted sync with multi-device passports. Every hop carries a receipt; the local daemon stays the source of truth.
Engineering notes, for the sceptical
- sync outbox Sync rides a transactional outbox: contributions land in a local append-only outbox that survives restarts and network failures, then deliver at-least-once. Kill the process mid-sync and nothing is lost.
- backpressure A full store refuses loudly with a typed error surfaced as a 429. A shared brain that silently dropped writes under load would be quietly forgetting; this one says no instead.
- day one Retrieval detects an under-populated graph and shifts its weight to lexical search, so a young fleet brain gives clean answers instead of noise from an empty index.
Measured, not vibes
Why shared memory beats bigger prompts
ScoreCrux, our open benchmark, tested the alternative: stuff the context window and hope. At production scale recall collapses while tool-mediated memory holds, and the token bill tells the same story. For a fleet, multiply both effects by every agent you run.
Recall at 2M-token scale
higher is betterFind 25 decisions + 5 needle facts across 3,346 documents.
28%
Stuffedmodel A · from 44%
8%
Stuffedmodel B · from 28%
80-100%
Tool-mediatedshared memory
More context makes strong models worse, at up to 6.6x the cost per run for far worse recall (ResearchCrux delta-cost-v1).
Context tokens, same answer
lower is betterNear-identical score; a fraction of the token bill, fleet-wide.
A fraction of the context tokens per agent for a near-identical answer, measured on the open ScoreCrux benchmark.
Methodology, negative controls and per-model runs: scorecrux.com/scale · scorecrux.com/context
One brain, governed
Shared memory needs shared rules
The moment five agents share one brain, governance stops being optional: who wrote this fact, who may read it, what did acting on it cost, and how do you unwind an agent that went wrong? Those are the control plane's three obligations (attribution, policy, revocation), and they ride the same receipts as the memory itself.
Design partners · limited intake · 2026 Q3
Run a fleet? Shape the control plane.
We are onboarding a small set of teams running real agent fleets. Two sentences is enough: fleet size, agent stack, and the governance question you can't answer today.