Memory for agents

Kaleidoscope

a kleos system · signed native binary

Scattered fragments of work, composed into one bounded view an agent can be handed. Runs on your machine, with adapters for 0xCopilot, Claude Code, Codex and Cursor.

A decision you make in one tool is invisible to the next one. Kaleidoscope is the store that sits underneath all of them, so the work you have already done stops being re-done.

Follow development ↗ Read the research ↗

Local

The store never leaves the machine.

Neutral

One store, many harnesses, no lock-in.

Bounded

You see exactly what was carried forward.

Planned adapters

4 harnesses

0xCopilot

planned

Claude Code

planned

Codex

planned

Cursor

planned

What we are building

01A local evidence store that captures decisions, corrections and conventions as you work — without shipping any of it off the machine.
02A context compiler that chooses what to carry forward before the model is called, and records why it chose it.
03Adapters, so the same store serves whichever harness you happen to open.
04An inspector, because a memory you cannot audit is a memory you cannot trust.

On evidence. Kaleidoscope is built on published research, and the benchmark that evaluates it is being built alongside it. We are not claiming performance numbers before that work runs — when there are results, they will appear on the lab site with the method attached.