Intelligence & Memory Systems

AI capabilities, knowledge management, and cognitive systems for Atom agents with 2026 Enhancement Plan integration.

📚 Quick Navigation

Start Here: Intelligence Overview - Complete intelligence documentation

Core Intelligence Systems

🧠 Memory & Learning

• Episodic Memory - Agent learning framework with POMDP (2026 enhanced) ✨
• Episodic Memory Quick Start - Get started with episodic memory
• Memory Integration - Memory system integration guide
• Context Memory (Per-Turn Fact Extraction) - Hermes-style durable-fact extraction; two-tier recall; extraction-first over compression-first ✨
• Atom vs. Hermes Comparison - Capability matrix + decision log vs. the Hermes agent
• Agent Memory Tools - memory_remember (INTERN+) / memory_forget (SUPERVISED+) — explicit agent-initiated persist/invalidate

✨ Self-Evolution (Enhanced 2026)

• Self-Evolution & Reflection Pool - Agents learn from mistakes and improve
  • Reflection Pool (critique-based storage)
  • Memento-Skills (generate skills from failures)
  • AlphaEvolver (optimize existing skills)
  • Arbor Framework Integration ✨ NEW: CodeHypothesisNode for skill evolution

🌐 Knowledge & Reasoning (Enhanced 2026)

• GraphRAG & Entity Types - Knowledge graph with multi-hop expansion ✨
• World Model & JIT Facts - Knowledge management
• JIT Fact Provision - Real-time fact verification with citations
• Vector Embeddings - Semantic search infrastructure

🐛 Debugging

• Debug Quick Start - Debug system setup
• Debug System - Debug architecture

🌳 Optimization (NEW 2026)

• Arbor Framework - Hypothesis Tree Refinement (HTR) ✨ NEW
  • CodeHypothesisNode for code generation optimization
  • WorkflowHypothesisNode for orchestration optimization
  • RoutingHypothesisNode for LLM routing optimization
  • POMDP integration for action space exploration

🚀 2026 Enhancement Plan Integration

All intelligence systems have been enhanced through the 2026 Enhancement Plan:

Phase 1: Enhanced Episodic Memory & Graduation ✅ COMPLETE

• POMDP Memory Framework: Write-manage-read loop for agent learning
• Memory Consolidation: Offline processing (inspired by human sleep)
• Experience-Driven Graduation: Quality-weighted episodes (20% improvement)

See: episodic-memory.md - Complete POMDP documentation

Phase 2: GraphRAG Enhancement ✅ COMPLETE

• Multi-Hop Expansion: Cue-driven activation for entity relationships
• Dynamic Graph Construction: Incremental updates (no full rebuild)
• Community Detection: Leiden algorithm for entity clustering

See: graphrag.md - Complete GraphRAG documentation with 2026 enhancements

Phase 3: Learning-Based LLM Routing ✅ COMPLETE

• RouteLLM Training: Preference data collection for router optimization
• Predictive Cache Warming: Pre-load frequently-used queries
• 15% Cost Reduction: Additional savings on top of existing cache

See: ../architecture/COGNITIVE_TIER_SYSTEM.md - Complete routing documentation

Arbor Framework Integration ✨ NEW

• POMDP Actions: Hypothesis nodes form action space in POMDP framework
• Observation Learning: Validation results feed into memory quality assessment
• GraphRAG Constraints: Failed hypotheses become negative constraints
• Routing Optimization: Arbor optimizes LLM selection for hypothesis generation

See: ../ARBOR_FRAMEWORK.md - Complete Arbor documentation

🎯 Key Features

Episodic Memory (Enhanced 2026)

• Hybrid Storage: PostgreSQL (hot) + LanceDB (cold)
• Retrieval Modes: Temporal (~10ms), Semantic (~50-100ms), Sequential, Contextual
• Graduation Integration: Episodes track agent learning progress
• Canvas Integration: Canvas presentations linked to episodes
• POMDP Framework: Formal write-manage-read loop with observation/action spaces
• Memory Consolidation: Offline processing for experience-driven learning

Self-Evolution (Enhanced 2026)

• Reflection Pool: Vector database of agent critiques
• Memento-Skills: Auto-generate skills from failure patterns
• AlphaEvolver: Optimize skills through mutation
• Maturity Gated: INTERN+ for critiques, SUPERVISED+ for optimization
• Arbor Integration: CodeHypothesisNode validates skill generation hypotheses

GraphRAG (Enhanced 2026)

• PostgreSQL-Backed: Recursive CTE traversal (<100ms)
• Multi-Hop Queries: Cue-driven activation for entity relationships
• Dynamic Construction: Incremental updates (no full rebuild required)
• Canonical Entities: 6 built-in types (user, workspace, team, task, ticket, formula)
• Custom Types: Dynamic entity type creation with JSON Schema
• Bidirectional Sync: Graph ↔ Database synchronization
• Community Detection: Leiden algorithm for entity clustering

Arbor Framework (NEW 2026)

• Tree-Based Refinement: Hypothesis Tree Refinement (HTR) for optimization
• Multi-Domain Support: Code, workflow, and routing hypothesis nodes
• MCTS Selection: UCB1 formula for exploration/exploitation balance
• Budget Enforcement: Tier-based limits (Free, Solo, Enterprise)
• Cumulative Learning: Cross-session negative constraints and insights

📖 Related Documentation

2026 Enhancements

• ARBOR_FRAMEWORK.md - Arbor Framework complete guide ✨ NEW
• ATOM_ENHANCEMENT_PLAN.md - 2026 Enhancement Plan
• VALIDATION_METRICS.md - Performance metrics

Agent Integration

• Agent Graduation - Promotion framework
• Auto-Dev Guide - Self-evolving agents
• Canvas Integration - Canvas presentations in memory

System Integration

• Cognitive Tier System - LLM routing optimization
• GraphRAG & Entity Types - Knowledge graph and entity extraction

🚀 Quick Start

# Episodic Memory with POMDP
from core.memory.pomdp_memory_framework import POMDPMemoryFramework
pomdp = POMDPMemoryFramework()

# Arbor Framework for optimization
from core.hypothesis_tree import HypothesisTree, OptimizationNode
tree = HypothesisTree(task_type="workflow")

# GraphRAG with multi-hop
from core.graphrag.multi_hop_expansion import MultiHopExpander
expander = MultiHopExpander()

# Self-Evolution with Arbor
from core.reflection_service import ReflectionService

---

Last Updated: June 18, 2026