Architecture and integrations

The desktop preview

The Windows and macOS packages use Electron. Its renderer is sandboxed, with context isolation and Node.js access disabled. A preload bridge exposes the desktop operations that the interface needs.

The setup core validates the chosen folder, prepares planned writes, creates skills and startup rules, checks the written files and stores a local installation profile. It records progress in a log.

Separate protocol, adapter and personal context

The universal protocol contains reusable memory behavior. Host adapters connect that behavior to the tools and permissions of a particular AI application. Personal preferences belong in a separate user layer.

Public documentation describes the shared mechanisms. Private notes, identity, relationships and conversational style do not belong in templates distributed to everyone.

Skills and startup rules

For Claude Code, setup writes a claudian-memory skill and a user rule. For Codex, it writes the skill under the shared agents skill directory and adds a marked block to the applicable global instruction file, with a backup of that file.

The selected vault path is written into the generated skill. This is why a later vault change needs coordinated integration updates rather than just a new path in the panel.

Claude Desktop can install a generated local Claudian Core extension. ChatGPT and Claude web use the hosted device relay: connect this device in Claudian, open the provider account connections, add the displayed address and approve the matching code. The computer and Claudian must remain running. Native account acceptance is separate from relay service tests.

The universal gateway direction

The broader architecture separates storage, the memory protocol, provider adapters and a context gateway. A gateway can enforce scoped reads and writes, preserve provenance and audit changes before they affect canonical state.

These are architecture goals and research components. The desktop preview’s local file setup should not be described as a completed universal gateway or provider OAuth connection.

The context panel

The embedded website panel has been withdrawn from the next desktop release while a dedicated companion interface is reconsidered. The local memory and connection manager remain available.

Selecting a local vault does not automatically upload it into that panel. The shared-memory installer and the web panel have separate data paths. A working context dashboard is not an always-on autonomous companion.

Runtime and model cost

Installing skills does not start an always-running reasoning agent. Memory work happens inside the AI application you already use and consumes that application’s available context and usage.

The cost of a future proactive system remains an open design problem. Event filtering and selective reasoning may help, but no zero-cost or always-on guarantee is made.