Dream Agent - Memory Consolidation System

The Dream Agent is SanctumOS's nightly, batch-orchestrated knowledge consolidation system. It operates over the day's conversations and historical archives to distill, integrate, and optimize knowledge.

Overview

The Dream Agent is a coordinated system of three neuro-modules that work together to maintain efficient and accurate long-term memory:

Hippocampus: Distills raw conversational activity into durable knowledge candidates
Neocortex: Integrates candidates into long-term knowledge structures
Glymphatic System: Compacts indexes and summaries, marking superseded beliefs

Architectural Summary

The Dream Agent operates as a nightly batch process that:

Processes Daily Activity: Reviews all conversations and interactions from the day
Extracts Knowledge: Identifies important facts, themes, and insights
Integrates Information: Incorporates new knowledge into existing structures
Optimizes Storage: Compacts and prunes to maintain efficiency

Design Tenets

Append-Only Truth

Nothing falls out of memory: Prior assertions are superseded, not deleted
Raw records remain immutable: Original data is never modified
Complete audit trail: Every change is tracked and reversible

Separation of Concerns

Summarization (Hippocampus): Focus on extracting knowledge from raw data
Integration (Neocortex): Focus on incorporating knowledge into structures
Pruning/Compaction (Glymphatic): Focus on maintaining efficiency

Deterministic Consolidation; Freeform Dreaming

Knowledge consolidation: Aims for stability and accuracy
Dreaming: Intentionally high-temperature and non-binding
Clear boundaries: Separate processes for different purposes

Provenance Everywhere

Source tracking: Every derived artifact traces back to source message IDs
Timestamp preservation: Maintains temporal context
Transform versions: Tracks how knowledge was processed

Budgeted Nightly Work

Incremental improvement: Few threads per night
Resource management: Controlled processing to avoid system overload
Quality over quantity: Focus on important knowledge extraction

System Context

Control Plane

The Dream Agent orchestrator that:

Schedules processing phases
Enforces resource budgets
Records metrics and performance
Coordinates between neuro-modules

Data Plane

The three neuro-modules acting over stores:

Hippocampus: Processing raw conversation data
Neocortex: Updating knowledge structures
Glymphatic: Optimizing storage and retrieval

External Actors

Broca/Thalamus: Finalize the day and reset daily buffers
SMCP/Letta: Provide IO (auth, fetch, write)
Direct Inference: Hippocampus runs direct inference to avoid agent-loop overhead

Module Responsibilities

Hippocampus - Consolidation

Responsibility: Convert canonical segments into knowledge candidates

Produces:

Layered summaries: Turn → topic → thread
Extracted facts: Subject-Predicate-Object + time + source + confidence
Themes: Recurring topics and patterns
Open loops: Unresolved questions and tasks
Per-segment embeddings: Vector representations for similarity

Processing:

Conversation Analysis: Review daily interactions
Fact Extraction: Identify verifiable statements
Theme Identification: Find recurring patterns
Context Preservation: Maintain temporal and relational context

Neocortex - Integration

Responsibility: Integrate knowledge candidates into long-term structures

Produces:

Dossiers: Comprehensive profiles of entities
Timelines: Chronological event sequences
Ledgers: Factual knowledge bases
Relationship Maps: Connections between entities

Processing:

Knowledge Fusion: Combine new and existing information
Conflict Resolution: Handle contradictory information
Structure Updates: Modify existing knowledge structures
Relationship Mapping: Identify and maintain connections

Glymphatic System - Optimization

Responsibility: Compact indexes and summaries, mark superseded beliefs

Produces:

Compacted indexes: Optimized search structures
Summary hierarchies: Multi-level knowledge summaries
Supersession markers: Track outdated information
Retrieval optimizations: Fast access patterns

Processing:

Index compaction: Reduce storage overhead
Summary generation: Create efficient representations
Supersession tracking: Mark outdated information
Performance optimization: Improve retrieval speed

Processing Workflow

Nightly Cycle

Preparation: Broca/Thalamus finalize daily data
Hippocampus Phase: Extract knowledge from daily activity
Neocortex Phase: Integrate knowledge into structures
Glymphatic Phase: Optimize storage and retrieval
Cleanup: Reset daily buffers and prepare for next day

Resource Management

Budget Enforcement: Limit processing time and resources
Priority Handling: Focus on most important knowledge
Incremental Processing: Avoid overwhelming the system
Quality Control: Ensure accuracy and consistency

Configuration

Environment Variables

DREAM_AGENT_ENABLED=true
DREAM_AGENT_SCHEDULE="0 2   "  # 2 AM daily
DREAM_AGENT_BUDGET_HOURS=2
DREAM_AGENT_MAX_THREADS=3
DREAM_AGENT_QUALITY_THRESHOLD=0.8

DREAM_AGENT_ENABLED=true DREAM_AGENT_SCHEDULE="0 2 " # 2 AM daily DREAM_AGENT_BUDGET_HOURS=2 DREAM_AGENT_MAX_THREADS=3 DREAM_AGENT_QUALITY_THRESHOLD=0.8

Processing Options

Schedule: When to run the nightly cycle
Budget: Maximum processing time and resources
Quality Threshold: Minimum confidence for knowledge inclusion
Thread Limits: Maximum concurrent processing threads

Monitoring and Metrics

Key Metrics

Processing Time: How long each phase takes
Knowledge Extracted: Number of facts and themes identified
Integration Success: How well new knowledge fits existing structures
Storage Efficiency: Compression and optimization results

Health Monitoring

Processing Success: Whether nightly cycles complete successfully
Knowledge Quality: Accuracy and relevance of extracted knowledge
Storage Growth: Rate of knowledge accumulation
Retrieval Performance: Speed and accuracy of knowledge access

Dream Loop (Optional)

High-Creativity Mode

Speculative insights: Generate creative associations
Dream Journal: Record imaginative connections
Non-binding: Does not write to core knowledge stores
Exploratory: Focus on novel and creative thinking

Configuration

Enable/Disable: Optional feature for creative processing
Temperature Settings: Control creativity level
Output Format: How to record dream insights
Integration Options: Whether to influence main processing

Integration Points

With Memory Systems

Read Access: Access to conversation history and context
Write Access: Update knowledge structures and indexes
Query Interface: Retrieve relevant information for processing

With Other Components

Broca: Receive finalized daily data
Thalamus: Access processed input streams
Conscious Mind: Provide consolidated knowledge for reasoning
Basal Ganglia: Support task execution with relevant knowledge

Performance Characteristics

Processing Speed

Hippocampus: ~1000 messages per minute
Neocortex: ~100 knowledge items per minute
Glymphatic: ~1000 index operations per minute

Storage Efficiency

Compression Ratio: 70-80% reduction in storage
Index Size: 10-15% of original data size
Retrieval Speed: < 100ms for most queries

Troubleshooting

Common Issues

Processing Timeouts: Adjust budget or optimize algorithms
Knowledge Conflicts: Review integration logic
Storage Growth: Increase compaction frequency
Quality Issues: Adjust confidence thresholds

Debug Tools

Processing Logs: Detailed execution traces
Knowledge Inspector: View extracted and integrated knowledge
Performance Profiler: Identify bottlenecks
Quality Metrics: Monitor knowledge accuracy

Future Enhancements

Planned Features

Advanced NLP: Better understanding of context and meaning
Multi-Modal Processing: Support for images, audio, and video
Real-Time Learning: Continuous knowledge updates
Enhanced Creativity: More sophisticated dream processing

Research Areas

Knowledge Representation: Better ways to store and organize information
Conflict Resolution: Improved handling of contradictory information
Efficiency Optimization: Faster processing and better compression
Quality Assurance: Better validation of extracted knowledge

Related Components

Hippocampus: Knowledge extraction and summarization
Neocortex: Knowledge integration and storage
Glymphatic System: Storage optimization and pruning
Memory Systems: Core storage and retrieval
Conscious Mind: Knowledge utilization and reasoning

Dream Agent - The memory consolidation system of SanctumOS, ensuring efficient and accurate long-term knowledge management.