Inside ZERO1 Core / Shared memory
Your processes, decisions, and know-how belong at the center of the swarm. Evidence Memory is the shared-context layer we are building into Core—so each specialist can work from the same approved understanding.
Start with the evidenceResearch behind the approach
Reusable workflows, retrieval design, and coordination each matter. These studies help us ask better questions—not promise universal gains.
Not ZERO1 performance. External benchmark findings, not customer outcomes or business ROI.
01 / Reuse
ICML · 2025 · Peer-reviewed12 percentage points between the reported methods.
Reported methods, not an isolated memory-only effect or business ROI.
WebArena · GPT-4 · Table 1
AWM reuses routines from earlier runs. The paper also reports a 15.0% accessibility-tree-only BrowserGym variant; the 23.5% reference uses a different observation setup.
02 / Remember
ICLR · 2025 · Peer-reviewed5 percentage points with expanded indexing keys.
One retrieval configuration, not a universal memory improvement.
LongMemEval-M · GPT-4o · round values · top-10 retrieval · Table 3
Extracted user facts are added to the indexing keys. Original records remain available in both conditions. This is a retrieval-design comparison, not memory versus no memory.
03 / Coordinate
Preprint · 2025 · v2Coordination helped on Finance Agent and harmed performance across multi-agent variants on PlanCraft.
Task-dependent findings, not evidence that more agents always help.
Diagrams: our design interpretation, not a live agent run.
180 configurations · four benchmarks · five architectures · three model families
The study compares architectures under standardized tool and token budgets. Splitting sequential work can add overhead. Our design inference is to choose the coordination pattern after understanding the task.
Reported by respondents in Microsoft’s global worker survey.
31,000 workers
31 markets
Our design responseGive agents the right company context. Then measure the result in the workflow where it matters.
What memory has to get right
Two recent benchmarks inform our design choices. These are controlled research settings, not production guarantees.
Design responseKeep original evidence attached to the answer.
Design responseRemember who said it—and what they can authorize.
The memory inside the swarm
Evidence Memory is part of Core’s Company Model. Follow one piece of information from its source to a result your team can inspect.
01 / Source
Start with approved records, policies, and prior work. Keep the source attached.
Proposed architecture, not a live data feed.
Illustrative policy record
A new policy should supersede the old one without erasing how earlier decisions were made.
Example schema. No customer records shown.
Learning, with your team
The learning loop we are building turns reviewed corrections into testable changes. Agents do not silently rewrite company rules.
Your team marks what should change.
Attach the correction to its source.
Check for gains and regressions.
Release a version that can be traced.
These are complementary tools, not mutually exclusive choices. Core’s design selects the retrieval method for the question.
The workflow and memory comparisons test particular methods. The coordination study finds gains and losses depending on the task. The survey and field study add workplace context. None evaluates ZERO1 Core.
The new percentage charts start at zero and end at 100. Differences are percentage points, not relative gains. The support-productivity bars use a derived index. Route F1 and answer accuracy are distinct measures. Authority-collapse squares represent tested configurations, not customers.
Our architecture responses are design inferences, not conclusions the researchers tested about our product.
The first working artifact is a synthetic RFQ-to-quote workflow. Founder-led diagnosis is current; live connectors, agent-led diagnosis, and governed production improvement are build work, scoped to each engagement.
The planned evaluation compares fixed tasks, model and tool budgets, source correctness, answer quality, permissions, latency, and cost. A sealed evaluation set and the comparative run are not complete. We are not publishing a ZERO1 benchmark score or a customer ROI claim.
See the first workflow2023 Work Trend Index. Vendor-funded survey; fieldwork February–March 2023. The 62% figure is respondents reporting search burden, not measured hours lost.
The Quarterly Journal of Economics, 140(2), 889–942 (2025). Peer-reviewed study. Reported average effects differ across worker groups.
arXiv:2607.25765v1, July 2026. Preprint. Displayed ranges apply to gold-constrained access across tested model backbones.
arXiv:2608.01679v2, August 2026. Preprint. Authority collapse means a stored claim loses constraints and acquires unwarranted instructional authority.
ICML 2025. Table 1; WebArena comparison.
ICLR 2025. Table 3; specified GPT-4o retrieval configuration.
Preprint, December 17, 2025 version (v2). Section 4.2; task-dependent coordination results.
