# @voooooogel — 2025-10-25

♥68 ↻8 · https://x.com/voooooogel/status/1981909701114921022

i've been working on an llm memory system testbed, where persistent kimi k2-based user simulators have conversations with transient models given access to a memory tool. i was curious what the effect of the boundary setting mentioned below was, so i let loose 35 kimi-simulated human spiritual seekers against three configurations:

- chatgpt-4o-latest, prompted with a system prompt very similar to the one used in chatgpt
- claude sonnet 4.5, prompted with a system prompt very similar to the one used in claude dot ai, including the boundary setting
- claude sonnet 4.5, using the same system prompt but with the boundary setting removed

in the image below, i've aggregated the results from each scenario, and scrambled them - they are not in the same order as listed above. i'm curious if people are able to guess which configuration A, B, and C are.

tags: author:voooooogel, kind:tweet, model:claude-sonnet-4-5, model:gpt-3-5, model:gpt-4o, model:kimi-k2, on:kimi-k2, year:2025
cited on: kimi-k2