$ cat ~/map/pi-taste.md
Pi Taste
Local preference-learning extension that distills accepted and rejected turns into one readable taste.md.
extensions✓ Sep 2
Facts
- repository
- github.com/LycanW/pi-taste ↗
- GitHub Stars
- ★ 2
- Forks
- 0
- license
- MIT
- npm / mo
- —
- last push
- Aug 31, 2026
install
pi install npm:pi-taste@0.5.6## TL;DR
A local preference-learning extension: a background Learner agent distills your accepted and rejected turns into a single readable taste.md — no cloud sync, no weight training — and injects it into future turns via <taste>.
## Key points
- $The Learner is a tool-calling model agent that runs after the foreground Agent settles; it only sees visible user and assistant text (thinking, tool results, and metadata are stripped) and edits taste.md itself via read_taste_file, write_taste_file, and edit_taste_file.
- $Automatic category reorganization kicks in once more than five learnings accumulate, splitting taste.md into per-category files; the full taste.md is injected on a future turn inside a <taste> tag.
- $The Learner follows the current main Pi model by default; /taste model points it at a separate one, and learning starts once the foreground Agent fully settles, possibly continuing in the background alongside later turns.
- $Taste is on by default with automatic learning and injection both enabled; /taste status checks state, /taste list reviews learnings, and /taste remember -g writes a global preference directly.
- $This release is tested with Pi 0.84.4 and requires Node.js 22.19 or newer; /reload picks up a fresh install or update.
## FAQ
Q Does it send my conversations anywhere?
No — the pipeline is local with no cloud sync; learning is done by the model agent already configured in Pi, and results land in a plain taste.md file.
Q Can I add a preference by hand?
Yes — /taste remember -g Always keep responses concise unless I request detail. writes a global learning directly, and /taste list shows everything learned so far.
## References
## Related entries
$ big-pi
Model-driven compression that summarizes completed phases into tiered blocks while Pi's JSONL transcript stays authoritative and untouched.
$ pi-agent-substrate
Suspendable remote workspaces for pi and Oh My Pi, on a self-hosted Agent Substrate cluster.
$ pi-ask-user-question
Adds a Claude Code-compatible ask_user_question tool that lets the LLM pop structured multiple-choice dialogs.
## Guides
Last updated: Sep 2, 2026