A portable operating layer for AI

One brain. Any model.

Keep your context, memory, tasks, history, and skills in one folder you own. Then let Claude, Codex, or the next useful model enter through its own door and work from the same approved truth.

The architecture

Different doors. The same brain.

Each tool can look for a different instruction filename. Those files should act as adapters into one canonical folder, not create a separate memory for every model.

Runtime A

Claude

Reads CLAUDE.md, then follows its routes into the shared files.

Owned layer

AI Brain folder

standing and current context
visible memory and decisions
open tasks and append-only history
canonical repeatable skills

Runtime B

Codex

Reads AGENTS.md, then follows its routes into those same shared files.

Another model arrives? Give it the entry file it expects and route that file to the same brain. Do not clone the brain per tool.

What makes it portable

Own the layer that compounds.

Plain files

Markdown and folders remain readable outside any one app. You can inspect, move, version, and back them up yourself.

One source of truth

Current state, history, tasks, and durable knowledge each get one home. Every runtime reads the same authoritative files.

Visible memory

Tool-specific hidden memory can help, but it is never the only copy. Important knowledge survives outside the runtime that captured it.

How it works

Connect, work, approve, repeat.

Connect

Give each runtime the instruction entry file it expects and point every entry to the same owned folder.

Load

The runtime reads only the smallest relevant slice of context, memory, tasks, and skills for the job.

Work

Use whichever model is best for the task. The model changes; the operating context underneath it does not.

Approve

Review proposed updates before they enter the brain. The next runtime starts from the new approved truth.

The portability test

Prove the brain survives the switch.

Start fresh sessions in two different runtimes from the same brain root.
Ask both to describe current priorities, open work, and the relevant working rules.
Approve one small update through the first runtime.
Confirm the second runtime can read the change from the shared files.
Fix any duplicate source or routing gap before adding automation.