Skip to content

v1.2.0: Dreaming Module — Self-Improvement for Any LLM

Latest

Choose a tag to compare

@Manzela Manzela released this 07 May 20:39
· 35 commits to main since this release

🧠 The Dreaming Module

Any LLM agent now gains a sleep cycle. The Dreaming Module analyzes past execution friction and proposes governance improvements — no fine-tuning, no RAG, no vector database. Just deterministic telemetry analysis transformed into actionable self-improvement.

What's New

  • ag-os dream — Full dream cycle: scans the Flight Recorder for friction (loops, rollbacks, budget overruns), synthesizes a Dream Report with root-cause diagnoses and proposed governance patches, and persists the report to ~/.antigravity/dreams/ as long-term memory.
  • ag-os dream --recall 5 — Retrieve past learnings from long-term memory.
  • ag-os demo --dream — The killer demo: simulates a failing, looping agent and then invokes the Dream Engine to diagnose and propose fixes.
  • MCP Tools: dream and recall_dreams expose the full Dreaming Module to AI agents via MCP.

Friction Detection (5 Archetypes)

Archetype Detection Severity
Loop Detected >2× max_loop_count transitions HIGH
Rollback Cycle ≥2 ROLLED_BACK visits HIGH
Budget Exceeded Solvency failure in metadata CRITICAL
Blocked Terminal Operation ends in BLOCKED MEDIUM
Excessive Transitions >10 transitions without COMPLETE MEDIUM

Try It

pip install ag-os==1.2.0
ag-os demo --dream

Quality

  • 59/59 tests passing
  • Zero lint errors (ruff)
  • Zero format violations

Full Changelog: v1.1.0...v1.2.0