Open source · verified knowledge · controlled execution

emmethAI

An open-source, local-first architecture for verified, reusable knowledge — with clearly separated responsibilities for knowledge, communication and execution authority.

VemmethAI is currently developed as a private open-source project. At the center is the CRN, surrounded by VOSS for authority, VASS for standardized agent integration, and the Agent Loop for controlled autonomous work.

Understand the CRN → Project status

Open-source core Local-first architecture Evidence before authority

01 / HEART

Core

CRN

Cognitive Retrieval Network. The CRN is emmethAI's knowledge layer: it preserves verified knowledge, orders evidence, and only returns knowledge when it fits the current context.

Knowledge, not transcript

The CRN is not a chat history and not a store for raw model output. It separates claim, evidence and admitted knowledge.

Context-bound reuse

Stored knowledge is never injected blindly. Relevance, applicability and conflicts are checked before knowledge is provided to a new task.

Conflicts become visible

Contradictory claims are not smoothed over by majority logic. Uncertainty can lead to abstain instead of false authority.

ObserveResult / error

VerifyCheck evidence

DistillExtract mechanism

AdmitAdmit knowledge

ReuseReuse when fitting

02 / AUTHORITY

Kernel & standard

VOSS / VASS

VOSS and VASS have different jobs. This separation prevents intelligence from automatically becoming authority.

VOSS Kernel

emmethAI Orbital Systems Standard. VOSS is the deterministic authority kernel. It does not decide what sounds creatively good, but whether a concrete state change is actually allowed to happen.

Execution and permissions are controlled.
Uncertainty can be handled fail-closed.
Models receive no independent authority.
Decisions can be traced and evidenced.

VASS Standard

emmethAI Agent Systems Standard. VASS describes the standardized connection between agents, models, tools, evidence and authority — without merging these components.

Typed requests instead of vague free-text interfaces.
Evidence and results stay distinguishable.
Models, tools and validators stay interchangeable.
VOSS remains the separate authority owner.

03 / AUTONOMY

Next layer

Agent Loop

The Agent Loop connects goals, planning, tools, verification and learning into a controlled work cycle. It is already part of the private development program, but there is no public release date.

Autonomous execution layer

In private development

The loop is designed to do more than answer tasks: it checks existing knowledge, plans steps, uses tools, verifies results, treats failures as new work items, and feeds verified findings back into the CRN.

AuditUse what exists

PlanDetermine steps

ExecuteApply tools

VerifyCheck result

LearnDistill verified

● Private development — no public release date

LOOP

04 / NETWORK

Planned beta

CRN Server

The planned CRN Server beta connects local CRN clients without turning private projects into a central dataset. Verified improvements flow to clients; client feedback is limited to abstract error classes.

Project context stays local

Project files, source code, raw prompts and private working context remain on the local system. The shared server path is not designed to require that material.

Verified updates within 24 hours

During the beta, connected clients are planned to receive newly admitted CRN improvements at least once within each 24-hour cycle.

Monthly public GitHub snapshot

Independent of beta access, the public CRN state is planned to be published as a versioned GitHub update once per month.

LocalError occurs

ClassifyAbstract error class

VerifyCheck mechanism

AdmitAccept improvement

SyncDistribute to clients

The CRN Server is not publicly available yet. The beta is planned as a controlled test of synchronization, privacy boundaries, conflict handling and distribution under real operating conditions.

05 / STATUS

Project status

Open source. Power for everyone.

VemmethAI is currently a privately developed open-source project. The open-source core will remain; optional commercial functions and services may be added later as an additional layer.

Today CURRENT

Development, testing, infrastructure and research are currently carried out privately. There is no public CRN Server access and no commercial product offering today.

Private development project
Open-source technical core
CRN Server beta still planned
Public releases only when evidence supports the claim

Later FUTURE

Future commercial capabilities may add hosted infrastructure, management features or professional services. They are intended to complement the open-source project, not replace it.

Open-source core remains available
Optional commercial services may be added
Local operation remains a central design goal
No promise of dates before the technology is ready

● Current state — private development

06 / TRUST

Privacy & transparency

What we protect — and what we make visible.

Transparency does not mean exposing trade secrets or security details. We publicly explain the principles, data classes and decision boundaries of the architecture.

Data minimization by designVemmethAI is designed so private project context stays local. Cross-project learning is limited to abstract mechanisms and error classes that do not require the underlying project content.

Allowlist instead of retroactive redactionFor global contributions we define which fields may be transmitted at all — instead of collecting as much as possible and cleaning up later.

No secrets as knowledgePasswords, tokens, private keys, seed phrases and comparable credentials belong neither in the CRN nor in public telemetry.

Evidence before knowledgeA model output does not become knowledge just because a model produced it. Verification and admission are separate steps.

Minimal necessary telemetryObservability should make system states traceable without unnecessarily mirroring raw prompts, source texts or sensitive content.

Authority stays separateAn intelligent model must not decide on its own which privileged action is executed. That boundary belongs to the kernel.

Publicly explainable — without exposing internal implementation

We explain: We publicly document the fundamental architecture roles, data classes, privacy principles, authority boundaries, and the intended behaviour in cases of error, uncertainty and conflict.
We do not publish: In particular, the following are not made public: secrets and key material, internal network details, security-critical implementation details, concrete exploit paths, proprietary heuristics, and sensitive operational and infrastructure parameters.
CRN contributions: During the planned CRN Server beta, client-to-server feedback is limited to abstract error classes. Project files, source code, raw prompts and private working context are not part of that feedback channel.
Conflicts: Contradictory information is not decided automatically by majority, popularity or the number of its producers. A conflict can therefore deliberately lead to ABSTAIN or to separate conflict handling.
Law & standards: Regulatory and normative information should be linked to its source, scope, jurisdiction and point in time. Changes must remain traceable and relevant information must be regularly checked for currency.
Certifications: VemmethAI does not claim any certification, conformity or legal property that has not actually been verified and evidenced. Terms such as ISO certification, GDPR conformity or comparable compliance statements are only used when the corresponding requirements are formally met and demonstrable.

This page describes the technical privacy, trust and transparency principles of the VemmethAI architecture. The legally required website privacy policy is considered separately and depends on the services and processing operations actually used — for example hosting, server logs, contact options, cookies, analytics or external services.

07 / SUPPORT

Independent development

Help move VemmethAI local.

VemmethAI is developed by one person and is currently fully self-funded. API usage, local hardware and CRN server infrastructure are paid privately.

One developer

Architecture, implementation, testing, documentation and infrastructure are currently developed and maintained independently by a single developer.

22B

≈ 22 billion API tokens / month

The current research and development workload uses roughly 22 billion model API tokens per month. The long-term goal is to move as much of that work as possible onto local hardware.

LOCAL

More local compute

More capable GPUs, more VRAM, memory and dedicated systems reduce external API dependency and make development more reproducible and self-controlled.

What support helps fund

Local AI hardware and larger-memory GPUs
CRN server, storage, networking and backups
API costs during the transition to local inference
Benchmarking, verification and open-source development

The goal

Move VemmethAI step by step from an API-heavy development workflow toward a predominantly local, independently operated research and execution environment — while keeping the open-source core available.

Support VemmethAI

VemmethAI is currently a private project and not a charitable organization. Support is voluntary, is not automatically tax-deductible, and no donation receipts can currently be issued.

emmethAI

Open source. Verified knowledge. Local-first.

VemmethAI is being developed privately and released step by step. Follow the architecture, the planned CRN beta and the path toward more local AI infrastructure.

Contact View architecture

Contact: [email protected] · Impressum · Datenschutz