إنتقل إلى المحتوى الرئيسي

Masar

Masar is a world model service for AI agents. It provides three capabilities that LLMs lack: structured planning, fast verification, and experience-based memory.

Why Masar?​

LLMs are good at generating text and code. They are bad at knowing what to generate, in what order, and whether the result is correct. An LLM asked to build a complex schema will over-engineer, under-build, or generate things in the wrong order.

Masar solves this by acting as System 2 for your agent:

  • Planning: "Here are 14 instructions across 7 dependency levels. Execute them in order."
  • Verification: "This result has a 94% chance of being valid. Predicted errors: none."
  • Memory: "You've handled 15 similar tasks before. The pattern that worked: check the KB first."

Three API Surfaces​

Planning API​

Given a partial result and a goal, produces dependency-ordered instructions:

plan = client.plan(current=partial, goal=target)
# Returns: 14 instructions across 7 levels
# Level 1: Add structure
# Level 2: Add states and events
# Level 3: Add transitions
# Level 4-7: Effects, UI, wiring, pages

Verification API​

Predicts validity and error categories without running the full compiler:

check = client.verify(schema=result)
# Returns: { valid: true, probability: 0.94, errors: [] }

Memory API​

Stores, recalls, compresses, and forgets agent experiences:

# Store
client.memory.store(episode=completed, domain="helpdesk")

# Recall
similar = client.memory.recall(context="database timeout", domain="helpdesk")
# Returns: matched pattern, 5 similar past episodes, common resolution

Next Steps​