AI Agents on an Infinite Canvas
Every node on your map is a full AI agent. It can search the web, analyze files, generate interactive charts and diagrams, and build on context from every ancestor node. Branch, instruct, compress, merge, and watch your research unfold as a living graph.
Research map with instruction, memory, and research nodes, connected by context inheritance
Each Node is an AI Agent
This is not a simple chatbot. Every node on your canvas is an autonomous research agent that inherits context from its ancestors and has access to real tools.
Real-Time Web Search
Each agent searches the internet for up-to-date information: market data, papers, docs, news.
Interactive Artifacts
Charts, Mermaid diagrams, SVG graphics, and full HTML widgets render right inside the node.
File Analysis
Upload PDFs, spreadsheets, code, or data. The agent analyzes content; files are inherited by descendants.
Full Context Awareness
Every agent sees the complete history from all ancestor nodes. Branch deep and context is never lost.
How Context Flows
Your research builds like a conversation tree. When you branch from a node, the new agent inherits everything its ancestors discussed: files, instructions, insights.
Branch
Split into sub-topics. Each branch gets its own agent with full ancestor context.
Inherit
Child nodes see the entire history from all parents. Context flows automatically.
Merge
Connect any two nodes manually. Multiple parents = merged context from different branches.
Node Types
Four roles: research agents carry the work; instruction, memory, and killed shape how context flows and what shows up in exports.
Research
The default agent on your map
The first card on a map is a research node, and every branch adds another—they are the same kind. Each one is an independent agent with its own thread, full ancestor context, web search, file analysis, and interactive charts, diagrams, and code.
Instruction
Set rules that propagate to every descendant
Like a team lead's brief that every team member follows. Phrases like "always include code examples" or "respond in Spanish" flow through the entire sub-tree automatically.
Memory
Compress hours of research into a brief
When a branch grows long, create a memory node to distill the key insights. Descendants see the concise brief instead of the raw conversation, so context stays focused and token-efficient.
Killed
Keep dead-end branches out of downstream context
Not every research path pays off. Mark a node as killed to dim it on the map and exclude it from exports. Killed nodes are omitted when building context for descendants, so irrelevant threads no longer shape what child agents see. Toggle with the K key.
Use Cases
See how different teams use Bentry. Each example shows a real graph structure with nodes, instructions, memory, and context inheritance.
Patterns to try
- Instruction at the top: set the research style early so every descendant follows it
- Memory before branching: compress a long thread before exploring new directions
- Multiple instructions: combine constraints for precise control
- Memory merge: synthesize insights from parallel branches into one node
- Kill dead ends: archive branches that didn't work out
Getting Started
Ask a question
Type your research topic and press Enter
AI researches
The agent streams a response with web search
Branch out
Explore sub-topics as new branches
Build the graph
Add instructions, memory, merge branches
How to branch
Click a node to open the chat panel. Type your message.
Click Branch & Send (fork icon) or press ⇧⌘↵
A new node is created. The AI responds with full ancestor context.
Manual connections: Drag from any node's bottom handle to another's top handle to merge context from multiple branches.
Capabilities
Web search
Agents automatically search the internet when they need up-to-date information.
Page fetching
Share a URL and the agent downloads and analyzes page content directly.
Parallel streams
Multiple nodes can research simultaneously and explore different angles at once.
Rich artifacts
Charts, diagrams, SVG graphics, and interactive HTML widgets, all rendered inline.
Supported file formats
.docx
Word
.xlsx
Excel
.csv
CSV
.json
JSON
.py .js .ts
Code
.md .txt
Text
.html
HTML
Up to 5 files per message, 20 MB each. Paste arXiv URLs to auto-download papers.
Markdown export
Export your entire map as structured Markdown. Killed nodes excluded.
Auto-layout
Press ⌘L to automatically arrange your graph.
arXiv integration
Paste an arXiv URL; the PDF is fetched and attached.