Memory model

Memory Model¶

Sonality memory has two persistent layers and one narrative layer:

• Graph memory (Neo4j): causal and temporal structure
• Vector memory (PostgreSQL + pgvector): semantic retrieval
• Personality state (SpongeState): compact evolving narrative and beliefs

Graph Memory¶

Main node types:

• Episode
• Derivative
• Topic
• Belief
• Segment
• Summary

Main edges:

• DERIVED_FROM
• TEMPORAL_NEXT
• DISCUSSES
• SUPPORTS_BELIEF
• CONTRADICTS_BELIEF
• BELONGS_TO_SEGMENT
• CONSOLIDATES

This layer supports provenance, temporal traversal, and belief-aware retrieval.

Vector Memory¶

derivatives table stores derivative embeddings for fast semantic search.

semantic_features table stores extracted long-lived features:

• personality
• preferences
• knowledge
• relationships

This layer supports retrieval quality and context enrichment.

Personality State¶

SpongeState stores:

• narrative snapshot
• topic opinion vectors
• belief metadata (confidence/uncertainty/provenance)
• staged updates
• pending insights
• recent shifts

Belief confidence follows LLM-assessed uncertainty during updates; legacy evidence-count confidence formulas are removed.

Consolidation and Forgetting¶

• Consolidation creates summaries from eligible segments.
• Forgetting archives low-value raw episodes while preserving recoverable graph structure and metadata.

Both operations are LLM-guided and run in reflection cycles.