# Eko memory carries skills and errors

> Eko memory helps and hinders.

By BIGCHANGE.AI Editorial

Published: 2026-10-25T06:06:39.341Z
Updated: 2026-10-25T06:06:39.341Z
Canonical: https://bigchange.ai/blog/eko-persistent-text-memory-island-task-results

![Conceptual charcoal illustration of a stepped virtual island with a rock, loose cubes and a short cube stack below a taller platform.](https://bigchange.ai/api/media/file/eko-island-rock-blocks-hero-v1.png)
AI-generated conceptual illustration by BIG CHANGE; no actual Eko trajectory or task result is depicted.

Eko, an Opus 4.7 coding agent in an [October 5 preprint](https://arxiv.org/html/2610.07130v1), kept files through hourly session and simulated island resets. Four tasks compared thirteen 30-hour histories with thirteen fresh agents.

## The big change

- **What changed:** Text preserved techniques across resets; weights stayed fixed.
- **Why it matters:** Builders can isolate useful notes; false notes can hurt tasks.
- **What to watch:** Validate stored rules before reuse.

## Evidence from tasks and memory edits

Experienced agents built more towers and placed more on the Spire, a high platform. More threw a block onto it without climbing. Tallest-tower scores stayed similar.

Three histories had five intact and five edited tests per goal. Erasing rock-throw notes about extra blocks reduced towers. Removing a false belief that placing could not reliably stack blocks improved another tower; a third edit slowed throwing but helped placement. Repeats were not independent histories. Physical transfer and subjective experience remain untested.

The [Clawblox repository](https://github.com/nacloos/clawblox) contains the public engine and mesa-island world. The [supplementary dataset](https://huggingface.co/datasets/nacloos/machine-playing-paper) lists hourly recordings, session logs and workspace snapshots; each `agentXX` is one trajectory. The authors’ [replay site](https://machineplaying.org/) shows the world and interface. These artifacts support inspection, although BIG CHANGE did not run the code or recompute the scores.

## Sources

- [Is this machine playing?](https://arxiv.org/html/2610.07130v1) — Author-reported methods and outcomes.
- [Clawblox source repository](https://github.com/nacloos/clawblox) — Authors' public engine and mesa-island world; repository presence and README checked, but code was not run or audited.
- [Machine-playing supplementary data](https://huggingface.co/datasets/nacloos/machine-playing-paper) — Authors' trajectory, recording, session-log and hourly-workspace listing; dataset was not downloaded or independently scored.
- [Machine Playing project and replays](https://machineplaying.org/) — Authors' explanatory replay and interface site, useful for inspecting presentation of the simulator; not independent evidence.