Parnee

Ignored
Rejected
Doubted
Will Succ.... Never Give Up

[Parnee @Parnee0wl]

Repo: https://t.co/98BLjL3l5E
Issues: https://t.co/aDi30dISIq

Or paste the README's setup prompt into your agent: "Read https://t.co/HYg8C9pYza then set me up on Beacon" (pick Local if asked).

Not install-tested by us; facts from the README.

Which agents do you switch between?

[Parnee @Parnee0wl]

Hit your Claude Code limit, switch to Codex, then spend 20 minutes re-explaining your project?

Beacon is open-source (MIT) memory shared across Claude Code, Cursor, Codex, OpenCode and 20+ other agents.

What it does, safe setup, open issues and a handoff prompt 🧵

[Parnee @Parnee0wl]

[Prompt] 3/3

4. VERIFY: run the listed test command and tell me if the result matches the memory. Don't edit code until I confirm the summary.

[Parnee @Parnee0wl]

Paper: https://t.co/A7Z9WRTLSC
Code: https://t.co/FTqYFBWzqR

Google Cloud AI Research + Northwestern.

What's the one line in your agent's instructions that changed its behavior the most?

[Parnee @Parnee0wl]

Same frozen model. Up to +11.5 points on an office-work benchmark.

Google didn't touch the weights. It had an AI editor rewrite the agent's harness, task by task.

Harness² in 6 lines 🧵

[Parnee @Parnee0wl]

(4/4)

4. VERIFY: tell me how to rerun the same task with A, with B and with both, and what difference in the output would prove each one worked. I keep only what the rerun supports.