Prompt Tree

English · 中文

Prompt Tree is a local-first AI chat client that models each conversation as a node-based DAG (a tree with branches) and provides an explicit Context Box for prompt assembly.

Motivation

Most chat UIs are linear:

• Exploring alternatives means copy/paste, losing context, or starting over.
• Long threads are hard to navigate and expensive in tokens.
• “Context management” is implicit, fragile, and hard to audit.

Prompt Tree turns a conversation into a canvas you can branch, compress, and curate.

Key Features

• Conversation DAG (tree + branches): continue from any node and compare outcomes.
• Canvas + chat in one place: visualize the thread while you talk.
• Context Box: drag nodes into a curated context, reorder, preview, and track tokens.
• Long-term memory (3-layer):
  • User Profile (JSON → derived Markdown): persistent preferences/identity, fully visible & editable in Settings.
  • Memory Bank (items): atomic memories with tags, status, confidence, and sources; supports user/folder scope.
  • Folder Doc (JSON → derived Markdown): per-folder long-term doc, visible & editable in the folder page.
  • First-message base injection + per-message memory refresh: on a thread’s first user message, inject Profile/Folder Doc + initial topK(folder) + topK(user) memories; on every user message, refresh auto memory hits in the Context Box.
  • Recent messages continuity (optional): on a thread’s first user message, copy the last N messages from your most recent thread (same folder) into the Context Box.
  • Async memory writer: runs on every user message; reads system prompt + all USER messages + a live memory snapshot resolved at execution time; can patch Profile/Folder Doc and upsert memories (first message can be forced to write at least 1 memory).
  • search_memory tool + pin: the model can search the full memory library during the tool loop; retrieved memories are added into the Context Box and can be pinned; per-thread caps are enforced (pinned wins over auto).
  • Embeddings (optional): embedding model is configurable; lexical fallback is used when embeddings are disabled/unavailable.
• Multi-model branching: select multiple models and send once to get parallel branches.
• Compression: compress a selected chain (or the Context Box) into a compact node; decompress when needed.
• Tools (optional): Web Search (Tavily/Exa), MCP servers, local Python execution.
• Local-first storage: conversations in IndexedDB; settings (providers/tools) in localStorage.

Install

Prerequisites

• Node.js 18+

Install (npm / pnpm)

npm i -g @yxp934/prompt-tree
# or
pnpm add -g @yxp934/prompt-tree

Update

npm i -g @yxp934/prompt-tree@latest
# or
pnpm up -g @yxp934/prompt-tree

Uninstall

npm rm -g @yxp934/prompt-tree
# or
pnpm rm -g @yxp934/prompt-tree

Run

Start the app:

tree

Open: http://localhost:1666

Change the port

tree --port 7777

CLI help / version

tree --help
tree --version

Run without installing globally (optional)

Useful if your system already has a tree command:

npx -y --package @yxp934/prompt-tree tree
# or
pnpm dlx --package @yxp934/prompt-tree tree

Usage Guide

1) Create folders & threads

• Create a folder (Library) to group threads.
• Optionally set a Unified System Prompt for the folder (applies to every new thread in that folder).

2) Configure providers (Model Service)

In Settings → Model Service:

• Add a provider
• Add one or more API keys (set a primary key)
• Set Base URL (OpenAI-compatible)
• Fetch/enable models and test connection

3) Select default models

In Settings → Default Model:

• Pick one or more enabled models (sending one message creates parallel branches)
• Pick a compression model (for summaries/meta instructions)
• Pick a title model (for thread titles)

4) Chat + branches

• Send a message once; if multiple models are selected, you’ll get multiple assistant branches.
• Use the branch list to switch, or continue from any node.

5) Navigate on the canvas

• The canvas shows the node graph of the current thread.
• Common actions: continue from here, compress branch, decompress, edit node, delete subtree.

6) Context Box (prompt assembly)

• Open the Context panel from the top bar.
• Drag nodes from the canvas into the Context Box.
• Reorder items, preview the compiled context, and track token usage.
• Compress either the whole context or a selected chain (must be a single continuous path).

7) Long-term memory (Profile / Memory Bank / Folder Doc)

• Configure in Settings → Memory:
  • Toggle long-term memory, auto-inject on first message, auto-add recent messages (last N), and the search_memory tool.
  • When auto-add recent messages is enabled, a new thread’s first user message will copy the last N messages from your most recent thread (same folder) into the Context Box.
  • In the chat composer Tools menu, you can toggle search_memory per message (default enabled on new installs). When enabled, it appears in the Context Box as a Tool Block and is included in Context Preview.
  • Pick Memory Writer model and (optional) Embedding model.
  • Edit/view User Profile JSON and its derived Markdown.
  • Browse/edit/delete/restore memories; optionally re-embed memories when you change embedding model.
• In folder threads:
  • Configure the first-message RAG ratio (topKFolder/topKUser) in the folder page.
  • Edit/view Folder Doc JSON and manage folder-scoped memories.

8) Tools (Web Search / MCP / Python)

In Settings → Tools, configure:

• Web Search: Tavily / Exa API keys
• MCP: add server JSON + token
• Python: python command, timeouts, output limits

Privacy & Data (based on current code)

• Conversations are stored locally in your browser’s IndexedDB database: AIChatClientDB.
• Long-term memory artifacts (User Profile / Folder Docs / Memory Bank + optional embeddings) are stored locally in the same IndexedDB database.
• Providers, API keys, and tool settings are stored locally in localStorage (keys prefixed with prompt-tree.*).
• The app runs locally; when you generate responses or use tools, requests are sent only to the model/tool endpoints you configure (this repo contains no built-in analytics/telemetry code).
• If you enable Memory Writer or Embeddings, the app will call the configured model endpoints for those operations.
• If you enable the Python tool, it executes code by spawning a local Python process on your machine—enable it only if you understand the risk.

Development

cd ai-chat-client
npm install
npm run dev

Open: http://localhost:1666

Useful scripts:

npm run build
npm run start
npm run test
npm run lint
npm run typecheck

Docs

• Product: 需求文档
• Technical design: TECHNICAL_DESIGN
• API design: API_DESIGN
• Roadmap: ROADMAP

Changelog

0.6.8 (2026-02-09)

• Markdown rendering: added remark-math + rehype-katex + katex to support inline ($...$) and block ($$...$$) formulas.
• Typography: added explicit h1-h6 styles under .prose-prompt-tree so heading levels render with distinct sizes.
• Formula display: improved .katex-display overflow handling to keep long equations readable in chat messages.

0.6.7 (2026-02-06)

• Settings: Max Tokens now supports empty value persistence; leaving it blank no longer resets to 1024 after saving.
• Defaults: for new settings, chat Max Tokens is now unset by default.
• Request pipeline: when Max Tokens is unset, chat and agent routes omit max_tokens instead of applying a fixed fallback.

0.6.6 (2026-02-06)

• Default Model settings: add an adjustable Prompt optimizer model selector.
• Prompt optimization request routing: now prioritizes the configured optimizer model; falls back to selected chat model, then chat default model.

0.6.5 (2026-02-06)

• Chat input: add one-click prompt optimization with loading state, cancel, undo replace, empty-input disable, and anti-concurrency guard.
• Settings: add editable optimizer prompt (plain text, no variable placeholders) and a Smart Memory toggle.
• Smart Memory for optimizer: when enabled, optimization requests include user profile, folder profile (if in folder), top 10 RAG atomic memories, and context snapshot.
• Prompt optimizer default prompt: switch to English and enforce output language based on input language while keeping technical terms/code as needed.

0.6.4 (2026-02-06)

• Provider settings: fix Base URL input behavior so / is no longer removed while typing; normalization now happens on blur/submit checks.

0.6.3 (2026-02-06)

• Memory: refresh auto memory RAG blocks on every user message instead of first-message-only injection.
• Memory Writer: build memory snapshot at job execution time from latest memory state to reduce duplicate/conflicting writes.

0.6.2 (2026-02-06)

• Memory: add “Auto-add recent messages” (last N) on a thread’s first user message to keep continuity across threads (in parallel with long-term memory injection).

Contact

• Email: yxp934@outlook.com
• WeChat: WanguA8

License

MIT