voicetreelab/voicetree

What it solves

Voicetree addresses the chaos of managing multiple AI agents and the "context rot" that occurs when LLMs are overwhelmed by massive conversation histories. It solves the problem of manual agent coordination and the lack of shared memory between humans and AI, providing a spatial way to organize tasks, plans, and progress.

How it works

The system uses a graph-based "mindmap" where nodes are markdown files and edges are wikilinks. Agents (such as Claude Code, Gemini, or Codex) live within this graph. When an agent is spawned on a node, it receives the contents of that node as its task and gains access to surrounding nodes as context, using a configurable radius and semantic search via local embeddings.

Agents can recursively decompose tasks into subgraphs, spawning their own parallel sub-agents (which appear as native terminals) to work through dependency graphs. Humans and agents share the same API and memory graph, allowing the system to restructure itself at runtime.

Who it’s for

It is designed for developers and researchers building complex, human-in-the-loop (HITL) coding-agent systems who need to externalize working memory and maintain high-level oversight of agentic workflows.

Highlights

• Spatial Memory: Uses a graph layout to organize agents and tasks, mirroring mental models to reduce cognitive load.
• Shared Human-AI Memory: Both humans and agents interact with the same markdown-based hypergraph database.
• Recursive Decomposition: Agents can break down large tasks into smaller connected nodes and orchestrate sub-agents to solve them.
• Speech-to-Graph: Includes a voice mode to capture ideas hands-free and turn verbal thoughts into structured mindmaps.
• Agent-First API: A lightweight API and CLI (vt) for CRUD operations on agents, lifecycle hooks, and graph navigation.