sayane 1.0.14.post1

Local-first persona and context portability for cross-LLM collaboration

Sayane (紗綾音)

日本語版 README · Documentation · Roadmap

Sayane is a local-first toolkit for carrying persona, context, prompts, and audit trails across LLM workflows.

It keeps your canonical context on your own machine, compiles it for different LLM runtimes such as ChatGPT and Claude, and prevents captured changes from being merged blindly.

Your context should not disappear when you change models.

Why Sayane exists

AI work is increasingly spread across multiple tools: ChatGPT, Claude, Gemini, Cursor, local models, browser extensions, and knowledge bases. Each runtime has its own memory, prompt format, and assumptions. As a result, serious users often lose:

• the context behind an answer,
• the persona or working stance used to produce it,
• the history of what changed,
• the ability to review whether a captured insight should become part of the canonical profile.

Sayane separates your context from any single vendor memory or chat service. It treats persona and context as local, reviewable project assets.

What Sayane does

Need	Sayane mechanism
Keep a canonical user profile locally	Sayane Profile in ~/.sayane/
Generate prompts for different LLMs	Profile → Prompt IR → target adapter
Avoid blind context merges	capture → candidate → evaluate → approve/reject
Track accepted and rejected changes	lineage records
Use local Markdown context	context/, storage index, optional Obsidian/Git workflows
Connect tools and editors	CLI, Local Bridge, MCP server, Chrome Extension

Current status

Community Edition in this repository is now v1.0.13 and supports the core local-first workflow.

Community Edition v1.0.13 is now published on PyPI.

Interface	Status	Primary use
CLI	Available	init / compile / candidate / storage
Local Bridge	Available	sayane serve + local HTTP API
MCP Server	Available	Cursor / Claude Desktop integration
Chrome Extension	Frozen / deprecated	legacy capture / context insert / candidate actions during migration
RDE/Candidate evaluation	Available	evaluate / approve / reject / lineage
Storage	Available	local Markdown / filesystem-first workflows, with Obsidian/Git kept as legacy compatibility paths

Try it in 5 minutes

Install the CLI first. See docs/install.md for detailed installation notes.

# PyPI (macOS / Linux / Windows with Python 3.11+)
pip install "sayane==1.0.13"

# macOS / Linux install script
curl -fsSL https://raw.githubusercontent.com/zyx-corporation/sayane/main/scripts/install.sh | bash

# Windows PowerShell install script
irm https://raw.githubusercontent.com/zyx-corporation/sayane/main/scripts/install.ps1 | iex

Verify the core flow:

sayane --version
sayane init
sayane compile --target chatgpt --profile examples/profiles/minimal.yaml

This verifies:

• local profile store initialization,
• target-specific prompt compilation from one profile,
• the basic flow: Profile → Prompt IR → Adapter output.

Notes:

• The Chrome Extension remains available for existing users, but it is frozen / deprecated and no longer the primary entry point.
• The recommended stable integration surfaces are the CLI, Local Bridge, and MCP Server.

Core workflow

Sayane Profile (local canonical context)
        ↓
Prompt IR (runtime-independent intermediate representation)
        ↓
Adapter (ChatGPT / Claude / Gemini / local targets)
        ↓
Target-specific prompt output

captured context → Candidate → evaluation → approve/reject → lineage

The LLM is not the owner of your persona. It is a runtime that receives a compiled prompt from your local profile.

Core principles

1. Separate persona from runtime

Your canonical persona and context belong to the local Sayane Profile, not to vendor memory.

2. Compile through Prompt IR

Do not copy one giant prompt string across runtimes. Compile one canonical profile into target-specific outputs.

3. Review meaning changes before merge

Captured context should become a candidate first. It should be evaluated, accepted, rejected, and recorded.

capture → candidate → evaluate (RDE/UIB-inspired review) → approve/reject → lineage

4. Prefer local-first portability

Sayane is designed so that serious AI work can remain inspectable, portable, and independent from any single hosted LLM memory.

Documentation

Start here:

• Installation
• Getting started (Japanese)
• CLI manual
• Local Bridge
• MCP Server
• Chrome Extension
• RDE / Candidate evaluation
• Storage / Obsidian / Git
• Docs index

Development

git clone https://github.com/zyx-corporation/sayane.git
cd sayane
uv run --extra dev pytest -q
# or keep a local venv explicitly:
# python -m venv .venv
# source .venv/bin/activate
# pip install -e ".[dev]"
# pytest -q

Development references:

• CI policy
• Development principles
• Roadmap

Positioning

Sayane is not a prompt snippet manager.

It is a small infrastructure layer for people who want AI collaboration to be:

• portable across models,
• local-first by default,
• reviewable before merge,
• inspectable after change,
• grounded in explicit persona, context, and lineage.

License

Apache License 2.0
SPDX-License-Identifier: Apache-2.0