How Hefty Thinks

Every message you send to Hefty flows through a series of steps that prepare Hefty to give you the best possible response. Here's what happens behind the scenes.

💬 Your Message
â–¼
🧠 Enrichment skills, knowledge, entities
parallel
📜 Context conversation history
â–¼
⚡ Reasoning & Execution think → plan → act → reflect loop
â–¼
📚 Learning extract & remember

Enrichment & Context

These two steps run in parallel the moment your message arrives:

  • Enrichment - Hefty searches its knowledge base for anything relevant to your request. This includes skills it has learned before, mistakes to avoid (antipatterns), and facts about your projects and tools (entities). All of this is fed into the reasoning step so Hefty draws on past experience.
  • Context gathering - Hefty assembles the relevant conversation history: recent messages, a summary of older exchanges, and any semantically similar past conversations. This ensures Hefty understands the full picture of what you've been working on.

Reasoning & Execution

This is where the real work happens. Hefty receives the full assembled context and decides how to help:

  • Simple questions - Hefty responds directly, just like a normal chat
  • Tasks that need action - Hefty creates a plan, executes each step using instruments, then reflects on the results

If a task requires multiple steps, Hefty works in loops - plan, act, reflect, and repeat until the job is done or it delivers a final answer. If something fails, Hefty sees the error and tries a different approach.

Everything streams to the UI in real-time - you see Hefty's thinking, plans, and action results as they happen.

Multimodal Support

Hefty can process images and files alongside text. You can attach files to your messages, and Hefty sends them to the LLM as multimodal content. When actions produce images (like screenshots), those are automatically fed back into the next reasoning loop so Hefty can see and react to visual results.

Learning

After each task, Hefty quietly analyzes what happened and extracts anything worth remembering:

  • New skills from techniques that worked well
  • Antipatterns from mistakes to avoid next time
  • Entities - people, tools, and concepts it encountered

This runs in the background and doesn't slow you down. Over time, it means Hefty gets better at helping you. See Learning for details.

Conversation Memory

As conversations grow longer, Hefty generates rolling summaries - compact narratives of older exchanges that preserve key details. This lets Hefty maintain context across long sessions without losing track of what you've discussed.