Your agents need a floor plan, not a chat window

What it does

Voicetree gives you a canvas of markdown nodes and edges that you inhabit alongside terminal-based agents like Claude Code or Codex. You place tasks, plans, and progress updates in a spatial layout; agents read nearby nodes for context and spawn child agents in their own subgraphs. Both humans and agents use the same lightweight API and in-memory hypergraph, so the system can literally rewrite its own structure while everyone stays oriented.

The interesting bit

The project treats the graph as a shared brain: agents retrieve context by proximity and semantic search rather than dumping entire chat histories, which the authors claim avoids the context rot that degrades LLM performance. Because agents are native terminals with full transparency, they can orchestrate parallel subagents and refactor their own workspace without turning into opaque black boxes.

Key highlights

• Humans and agents share identical API access to the markdown hypergraph, enabling genuine human-in-the-loop orchestration.
• Agents decompose tasks into subgraphs and spawn recursive child terminals, keeping progress visible rather than buried in chat logs.
• Context retrieval uses spatial proximity and local embeddings instead of monolithic conversation history.
• Includes a speech-to-graph mode for hands-free idea capture, turning voice into structured nodes.
• Self-described as “early-beta” and “powerful but rough”; the team dogfoods it daily for AI forecasting and research.

Caveats

• Explicitly labeled “early-beta” and “powerful but rough,” with development described as “spiky.”
• Source-available under BSL 1.1; full Apache 2.0 conversion is four years out.

Verdict

Worth a look if you are already running multiple terminal agents and need a shared memory surface that both you and they can refactor. Skip it if you want a finished IDE or immediate open-source licensing.