The Gap Corporate AI Can't Fill

Today's leading AI assistants are genuinely impressive, and all of them now offer some form of personal memory — a profile built from your conversations over time. But there's a fundamental difference between memory that lives on a corporation's servers, feeds into their models, and exists to improve their product, and memory that lives on your hardware, serves only you, and is never seen by anyone else. Corporate AI memory makes the corporation's AI better at serving everyone - not the individual. ChatGPT runs on a surveillance business model — your conversations are collected, analyzed, and used to improve a product that serves millions of other people. That's a requirement of their business - advertising is the only way that business model can turn a profit. As long as that model stands, privacy isn't something they can offer, because real privacy would put them out of business.

People discuss deeply personal things with AI assistants: health anxieties, grief, relationship struggles, financial stress, concerns they might not share with another person. That's not surprising — AI is available at 3am, non-judgmental, and often genuinely useful in difficult moments. But it's worth asking where that vulnerability goes. When it's stored as part of a profile on a corporate server, it becomes data. When it lives on your own device, under your control, it stays private.

What exactly happens to your data — your conversations — on a corporate server isn't always clear — and that opacity is by design. OpenAI, for instance, has entered a formal partnership with Palantir, a company whose core business is building surveillance and data analytics infrastructure for government and military clients. The details of what flows where under that arrangement aren't publicly disclosed. We don't know what these companies ultimately do with the conversations their users have in vulnerable moments, and trusting that it stays private isn't logical — it's wishful thinking. Zynkbot's memory makes it better at serving you specifically — and that knowledge goes nowhere.

That your most intimate thoughts should be available only to you is not a technical preference — it's an ethical position. It used to be an assumption so basic it didn't need to be stated. If a store clerk followed you around, writing down everything you said and noting every item you paused to look at — the way Amazon tracks exactly how long you hover over each product — then sold that information to companies wanting to steer your future decisions, you'd consider it a serious violation. We accept a near-identical arrangement from AI assistants and online businesses because it's framed as "improving the product." Over the past few decades, through the gradual normalization of data collection, behavioral profiling, and always-on connectivity, that assumption of privacy has been quietly dismantled — largely without opposition, and often without people noticing until it was already gone. Zynkbot doesn't solve that problem entirely, but it takes a clear stance on it, and gives you one practical place to draw a line.

Zynkbot isn't trying to replace powerful online AI models and tools, nor compete with them. It fills a different demand — one for the privacy they can't offer: a trustworthy companion that actually learns who you are over time, not to manipulate you, but to help you. That knowledge is stored privately on your own hardware and carried into every conversation — regardless of which underlying AI model you're using that day. The model is the face. The memory is the relationship.

Zynkbot didn't start as a response to government and corporate surveillance, although this has increased alarmingly in recent years on both fronts. It started with a more specific problem: building a private AI companion with a parenting mode — a way for children to have a phone where internet access is filtered through an AI, with parents staying in control until a child is ready to navigate it alone. But the same architecture that makes that possible — local memory, no cloud dependency, genuine user control — turns out to be exactly what everyone needs right now. The result is a trustworthy companion, but it also turns out to be a defensive anti-surveillance tool: a way to protect yourself in an environment that was quietly built to work against you.

The privacy focus isn't ideology — it's a practical response to a real shift. Over the past thirty years, the rise of social media and the quiet expansion of corporate and government data collection have fundamentally changed what it means to have privacy. Most people sense this even if they can't articulate it. If your most personal context — your goals, your health, your doubts, your half-formed ideas — is going to live somewhere digitally, it should live with you. Not harvested. Not used to train a commercial model. Not held on a server you don't control. Zynkbot is one practical response to that reality — a small piece of infrastructure that puts something back on your side of the equation.

🔒 Privacy in Practice

"Privacy-first" is easy to say. Here's specifically what it means in Zynkbot's design.

Local-First by Default

Zynkbot runs local AI models out of the box — standard .gguf format files compatible with llama.cpp. When running locally, nothing leaves your device. No API call, no network traffic, no external dependency. The inference happens on your CPU or GPU, the memory database is a local SQLite file, and the application itself is a Tauri desktop app with no telemetry or update pings.

Optional API Use — On Your Terms

If you want access to more capable models (GPT-4, Claude, Grok), you can connect your own API keys. When you do, your conversation text goes to that provider for inference — but your memory database never does. Zynkbot constructs the context window locally, retrieves relevant memories locally, and sends only the assembled prompt to the API. The underlying personal data stays home.

API providers are primarily business services. Their privacy policies are driven by enterprise compliance requirements and typically mandate prompt deletion within 30 days. By default, your conversations are not permanently retained and are not used to train their models. Zynkbot is designed to take full advantage of this — treating the API as a stateless inference service, not a memory store.

No Account. No Login.

Once Zynkbot is installed, it just opens. There's no account to create, no email verification, no password reset flow, no session that expires and locks you out. Your AI assistant is on your device — you don't need anyone's permission to use it.

No Telemetry. No Analytics. No Tracking.

There are no usage analytics, no error reporting services, no version check pings, no "help us improve" data collection. The application does not phone home in any form. You can verify this by inspecting the source code — it's all there under AGPL v3.

The License as a Privacy Guarantee

The AGPL v3 license isn't just an open source choice — it's a structural protection. AGPL requires that anyone who modifies and distributes Zynkbot must also release their modifications. This prevents a company from taking the codebase, adding surveillance features, and distributing a closed-source fork without disclosing what they've done. The license keeps the code auditable.

🛡️ Safety Containment Modes

Zynkbot includes a configurable safety layer — local TinyBERT classification runs by default on your device before content reaches any model. Modes are global settings for Zynkbot, as opposed to snap-ins which are industry-specific tools. Choose the containment level that fits your use case, or adjust/create modes for a specific use case.

The safety classifier runs entirely locally — no content is sent to an external moderation service. All modes can be changed at any time through Settings.

System Controls — switch containment (safety) modes and models from a single panel

🧩 Snap-ins Experimental

Snap-ins are purpose-built mini-applications that run inside Zynkbot for specific workflows. The therapist snap-in is a working demo of the concept — it indexes session notes by patient using RAG, making them fully searchable within the knowledge base. The current version is a proof of concept: indexing works, but querying through the interface is not yet implemented. Snap-ins are experimental and intended as a developer preview of what domain-specific tooling can look like on top of the Zynkbot platform.

In virtually every industry, workers could benefit from an AI with a snap-in focused on their field. In our personal lives, snap-ins can draw on data accumulated over time to track everything from mental and physical health to hobbies, games, and leisure activities.

Therapist snap-in — session notes organized by patient, indexed locally for RAG search. Fully private, stored on device. Have a conversation about a patient with your Zynkbot, and it will have complete history and detailed notes to draw upon - no corporate servers required.

Building on Snap-ins

Snap-ins are part of how this project sustains itself. ContainAI will develop industry-specific snap-ins — tools for lawyers, medical professionals, tradespeople, educators, and others — available for a monthly fee. Third-party developers are explicitly welcome to build and sell their own snap-ins on the same platform. The base Zynkbot app is and will remain free; the snap-in ecosystem is how the project grows without becoming dependent on advertising or data collection.

Optional Memory Backup

Zynkbot's memory database lives entirely on your device. An optional encrypted cloud backup service is planned — for a small monthly fee comparable to what you'd pay for cloud storage, your memory database can be synced to a secure server. Useful when you're away from your home network and need a device up to date, or as protection against device loss. Nothing about this is required; it's a convenience for people who want it.

⚡ ZynkCluster Upcoming

Most AI models are constrained by the VRAM of a single device. A model like Mixtral 8x22B requires roughly 140GB of VRAM — more than any single consumer GPU can hold. ZynkCluster proposes distributing that model across multiple devices on your local network, with each machine's GPU running a different part simultaneously. Models that are simply out of reach on a single consumer machine become accessible across hardware you already own — no cloud, no data leaving your network.

Zynkbot uses Mixture of Experts (MoE) models, where each query activates only a small subset of the model's expert sub-networks. The number of active experts per token depends on the model — Mixtral uses 2, DBRX uses 4, OLMoE uses 8. ZynkCluster maps those active experts to separate devices and runs them in parallel. Unlike existing distributed inference tools that pass every token through every device in sequence, ZynkCluster only involves the devices that are needed — and uses them at the same time.

ZynkCluster is built on the same device-pairing infrastructure as ZynkSync. The system that already syncs your memories across devices becomes the coordination layer for pooling their compute.

The approach is theoretically sound and grounded in published research on MoE architectures. Development is ongoing; implementation will proceed as other priorities, including Android support, are completed.