ZekeLabs

Kalsys

Status
In build
Role
Architect and operator. Specified who may write to memory and which actions an AI can take; directed Claude through the build and every review.
Stack
TypeScript, MCP, Postgres, Supabase Auth, Railway
Updated
Nov 2026

One memory shared across AI assistants. The assistant proposes a change, Kalsys checks it against written rules, and nothing is saved until the owner approves.

The Kalsys connect page: the headline Bring your memory to your AI client, a short description of how the assistant proposes, Kalsys checks and writes, and the owner approves every save, and a notice that the server is in a private build.

How it's put together

Each box is a part of the system. Dashed red marks the point where a person or a second model has to sign off.

AI assistantKalsys serverRule checkOwner approvalSaved memoryAI assistantKalsys serverRule checkOwner approvalSaved memory

Problem

AI assistants forget everything between chats, each tool keeps its own partial memory, and when they do save something they save mistakes: notes in the wrong shape, files the next session cannot read, stale copies that quietly win.

What got built

One server that holds the memory and owns every write to it. An MCP client (Model Context Protocol, the open standard for connecting AI models to tools) connects with a single address and an OAuth sign-in, so there is no key to paste. The private build accepts the owner account only. The assistant proposes a change; Kalsys checks it against written rules and shows the owner the exact plan; it is written only after approval. Work sessions are locked per project so two chats cannot write over each other.

Result

In daily use on ZekeLabs projects while still in build, with a test suite and a review by a second AI model at each stage. Not yet open to other accounts.

Want something like this built?

hello@zekelabs.dev