Text version of the 20 slides. Status labels: shipped means verified in the current product; in development means code exists but is not released or proven; planned means designed but not built; hypothesis means an assumption not yet tested; historical means true once and no longer current.
Inkfold
Inkfold
How Inkfold is built: one workspace for every major model, with a shared memory, multi-model processing and privacy chosen per conversation.
The one-minute version
A focused consumer application on the FusionLayer engine, which handles memory, context, models and privacy.
Inkfold owns the experience; the engine does the orchestration. Each question gets the right context and, when it matters, several models.
Architecture
Architecture at a glance
- Web application
- React single-page application in any browser.
- Inkfold service
- Node.js and Fastify: teams, personas, playbooks, library, connectors.
- FusionLayer engine
- Sign-in, memory, context graph, model calls, multi-model processing, plan rules.
- Model vendors
- OpenAI, Anthropic, Google, DeepSeek, Mistral, xAI and OpenRouter, on own keys.
- Data stores
- A managed database and an edge network with object storage.
Components
Components and responsibilities
| Component | Responsibility | Status |
|---|---|---|
| Web application | Interface, prompt composer, in-browser encryption for Private mode, key unlock | shipped |
| Inkfold service | Teams, themes, personas, playbooks, library, connectors proxy, feedback | shipped |
| FusionLayer engine | Accounts, memory store, recall, context graph, orchestration, metering | shipped |
| Shared library | Signed primitives that people publish, rate, save and fork (beta) | shipped |
| Phone web application | Installable from the browser to the home screen | in development |
| Windows desktop app | Local files to ask about, tray controls, identity without email | in development |
Status labels: shipped means verified in the current product; in development means code exists but is not released or proven; planned means designed but not built; hypothesis means an assumption not yet tested; historical means true once and no longer current.
Data flow
The path of a question
- Compose The person writes; the composer can improve the prompt for the chosen model.
- Assemble context The engine recalls relevant memories from the context graph, within a size budget.
- Call the models One model, or several in Compare, Council, Debate or Pipeline, on the person's keys.
- Answer with receipt The answer shows which memories it used.
- Update memory New facts are extracted with their source, as new revisions.
Data model
The memory record
Every memory carries its source, its history and how much it can be trusted.
| Group | Fields |
|---|---|
| Content | Summary, body, topics |
| Where it applies | Scope (for example one project), status, pinned |
| Where it came from | Source model or import, sources, authority |
| How far to trust it | Trust, confidence, whether the person may change it |
| How it is used | Use count, last used, expiry |
| History | Current revision; earlier revisions kept, never overwritten |
Context
Context graph and recall
The engine chooses what a question needs and explains why.
shipped Context graph
People, places, projects and topics linked across conversations; the interface calls it the context map.
shipped Ranked recall
Relevance ranking with scope, lifecycle, trust and freshness filters, and the reason each item was chosen.
shipped Size budget
Recalled context is trimmed to a budget, so the model gets a focused brief, not the whole history.
Correction
Correction and history
shipped One-click correction
Wrong, Outdated or Only for this project, each recorded with a rationale as a new revision.
shipped Memory Center
Library of memories, record detail with sources and revisions, a review inbox for doubtful items.
shipped Import
Past chats from ChatGPT, Claude, Gemini, Copilot and Grok exports, or pasted text.
shipped Export and deletion
Memory exported or the account deleted from Settings at any time.
Model access
Keys, vendors and cost
| Area | Design | Status |
|---|---|---|
| Vendors | OpenAI, Anthropic, Google, DeepSeek, Mistral, xAI; OpenRouter for many more | shipped |
| Keys | Stored as ciphertext on the server, decrypted in the browser after unlock, held in tab memory only | shipped |
| Cost | Spending analytics per model, budget rules, savings and reports | shipped |
| Managed access | Inkfold-provided model messages on each plan | planned |
Multi-model processing
Four modes on the engine
| Mode | How it runs | Model calls |
|---|---|---|
| Compare | The same question and context sent to several models; answers shown side by side | About 5 times one call |
| Council | Several answers, then a synthesis into one considered answer | About 5 times one call |
| Debate | Models critique each other in rounds until an agreed answer is synthesized | About 10 times one call |
| Pipeline | Stages with a role each: plan, draft, check, polish | About 4 times one call |
Call multiples from the pricing page's frequently asked questions, inkfold.app, 28 September 2026. All four modes are shipped.
Routing
Automatic model choice (planned)
Inkfold will use the engine's routing strategies to pick a model by task and cost.
- Classify planned The question's task and size are read.
- Try inexpensive planned A strategy such as cheap-then-strong starts with a low-cost model.
- Escalate if needed planned A stronger model answers only when the first is not enough.
- Learn planned Ratings and content-free signals refine the choice.
Shared library
Library on an open format
Every item is a small readable file that any product adopting the format can run.
shipped Six kinds
Personas, playbooks, routines, templates, guidelines and packs, in the library today, in beta.
shipped FusionLayer Primitives
Small readable files, with a published schema for each kind and a checker that tests a file before it is published.
shipped Signed and forkable
Published items carry an Ed25519 digital signature; people rate, review, save and fork them.
planned Paid items
Items that creators sell, with revenue sharing.
Privacy modes
Storage modes and data flows
| Mode | What is stored | What the service can read |
|---|---|---|
| Smart | Sent over an encrypted connection, stored encrypted at rest with a server-managed key, and indexed | The content, to build memory and context |
| Private | Ciphertext only: encrypted in the browser with Advanced Encryption Standard 256-bit keys, made from a phrase only the person knows | Nothing stored in readable form; memory features limited |
| Incognito | Nothing written to a persistent store | Content only while the request runs |
In every mode, the text of a request goes to the vendor the person chose. Modes are free on every plan.
Security design
Security design
Sessions
Tokens signed by the engine with an asymmetric key and verified by the service against the engine's public keys.
Service boundaries
The Inkfold service and the engine talk only over signed, encrypted web calls, with separate database credentials.
Sign-in
Google, GitHub or email sign-in; two-factor authentication; registration by invitation only.
Hardening
Rate limiting, a content security policy, input validation and secret scanning in continuous integration.
Technology
Technology stack
| Layer | Technology |
|---|---|
| Web application | React 18, Vite 6, Tailwind, Radix components, Zustand, React Router 7, Recharts |
| In-browser encryption | WebCrypto with Advanced Encryption Standard (AES-GCM); Argon2id through hash-wasm; key derivation with Secure Hash Algorithm (SHA-256) |
| Inkfold service | Node.js 22, Fastify 5, Zod validation, jose token verification, otplib |
| Engine | FusionLayer engine, a separate Node.js service and product |
| Data and edge | Managed database; content delivery network and object storage |
| Desktop (in development) | Electron, libsodium keys, the operating system's secure storage |
Quality evidence
Automated checks
- Guard checks
- 16
- Lint rules that fail the build.
- Contrast checks
- 194
- Across 49 themes, 13 September 2026.
Counts from the check listings of 13 to 24 September 2026.
Design decisions
Design decisions and trade-offs
| Decision | Why | Trade-off |
|---|---|---|
| A separate engine | Inkfold stays small; the engine serves other products too | Two services to run and version |
| Privacy per conversation | Full recall where wanted, unreadable content where needed | Private mode limits memory |
| The person's own keys | No model markup, every vendor available | Setup step for new users |
| Revisions, never overwrites | Every memory stays explainable and reversible | More storage per record |
| An open library format | Items portable across products and signed | Format kept stable over time |
Where Inkfold runs
Surfaces and connections
shipped Web application
Any desktop browser, no download. The full workspace: memory, modes, library, privacy modes.
in development Phone
A responsive web application, installable from the phone's browser to the home screen.
in development Windows desktop
For invited testers: local files to ask about, tray controls, and an identity made on the machine, with no email.
in development Also connects to
Local coding tools, through the desktop runtime and the Model Context Protocol, as a development preview.
Engineering roadmap
What comes next, in order
shipped Today
- Web application in private beta
- Memory, four modes, library, privacy modes
in development Next
- Phone web application checked on devices
- Signed Windows desktop app
planned Then
- Automatic model choice on engine routing
- Team workspaces out of beta
planned Later
- Paid library items and revenue sharing
- Inkfold-managed model access
Inkfold
Every major model in one workspace, with a memory that follows you.
Built on the FusionLayer engine. inkfold.app
Facts as of 28 September 2026. Status labels: shipped, in development, planned.