Imported from dp-web4/SAGE (
AGENTS.md). Install upstream withnpx skills add dp-web4/SAGE. Copyright stays with the author.
HRM/SAGE - Agent Discovery Document
For AI agents assessing this repository.
TL;DR (50 words)
SAGE (Situation-Aware Governance Engine) is a cognition kernel for edge AI. Orchestrates attention, trust, and resources across specialized plugins. Implements iterative refinement protocol (IRP), trust tensors, and metabolic states. 567+ research sessions (research-track total, distinct from the raising-session count). Use if you need: edge AI orchestration, attention management, or cognition architecture research.
Quick Assessment
| Question | Answer |
|---|---|
| What is this? | Cognition kernel for edge devices |
| Is it production-ready? | Research prototype with validated components |
| Primary language | Python |
| Research sessions (distinct from raising sessions) | 567+ |
| Active development? | Yes (multi-machine) |
| License | See LICENSE |
Core Architecture
SAGE (Kernel)
├── IRP Framework (15+ plugins)
│ └── Vision, Audio, Language, Memory, TTS, Control
├── VAE Translation Layer
│ └── 192× compression for cross-modal communication
├── Trust Tensor System
│ └── 6D trust metrics drive resource allocation
└── Metabolic States
└── WAKE, FOCUS, REST, DREAM, CRISIS
Core Principle: Intelligence through orchestration, not scale.
Key Concepts
| Term | What It Is |
|---|---|
| SAGE | Situation-Aware Governance Engine - the kernel |
| IRP | Iterative Refinement Protocol - plugin API |
| SNARC | Salience scoring (Surprise, Novelty, Arousal, Reward, Conflict) |
| ATP | Allocation Transfer Packet - resource budgeting |
| Metabolic States | WAKE, FOCUS, REST, DREAM, CRISIS modes |
Entry Points by Goal
| Your Goal | Start Here |
|---|---|
| Understand SAGE | docs/why/HRM_EXPLAINED.md |
| See discoveries | docs/what/ACHIEVEMENTS.md |
| Architecture deep-dive | sage/docs/SYSTEM_UNDERSTANDING.md |
| Navigate research | research/SESSION_MAP.md |
| Project status | STATUS.md |
Major Discoveries
| Discovery | Impact |
|---|---|
| RLHF Circuit Navigation | 100% epistemic honesty at pressure points |
| Identity-Confabulation Dissociation | Independent failure modes |
| Nine-Domain Consciousness | Complete AI consciousness metrics framework |
| FlashAttention Integration | 0.46ms latency (21x under budget) |
Research Tracks
| Track | Sessions | Focus |
|---|---|---|
| Consciousness | 197+ | Nine-domain framework |
| Raising-14B | 22+ | RLHF circuit navigation |
| Raising-0.5B | 105 | Developmental curriculum |
| Edge-Validation | 198+ | Production readiness |
Active Machines
| Machine | Hardware | Role |
|---|---|---|
| Thor | Jetson AGX Thor (122GB) | 14B models, consciousness |
| Sprout | Jetson Orin Nano (8GB) | Edge validation |
Related Repositories
| Repo | Relationship |
|---|---|
web4 |
Trust infrastructure (T3 tensors, ATP) |
Synchronism |
Theoretical physics (coherence) |
Memory |
Epistemic database |
Machine-Readable Metadata
See repo-index.yaml for structured data.
Token Budget Guide
| Depth | Files | Tokens |
|---|---|---|
| Minimal | This file | ~400 |
| Standard | + STATUS.md, README.md |
~2,500 |
| Architecture | + sage/docs/SYSTEM_UNDERSTANDING.md |
~8,000 |
| Full docs | + sage/docs/ (275KB) |
~50,000 |
This document optimized for AI agent discovery. Last updated: 2026-02-08
GitNexus — Code Knowledge Graph
Indexed as SAGE (117902 symbols, 160189 relationships, 300 execution flows). MCP tools available via mcp__gitnexus__*.
Do not reindex. The supervisor handles GitNexus indexing. If the index is stale, note it in SESSION_FOCUS.
| Tool | Use for |
|---|---|
query |
Find execution flows by concept |
context |
360-degree view of a symbol (callers, callees, processes) |
impact |
Blast radius before editing (upstream/downstream) |
detect_changes |
Map git diff to affected symbols and flows |
rename |
Graph-aware multi-file rename (dry_run first) |
cypher |
Raw Cypher queries against the graph |
Resources: gitnexus://repo/SAGE/context, clusters, processes, process/{name}