Designing a Knowledge Graph-Driven Personal AI Assistant Memory System

You are an expert in knowledge graphs, personal AI assistants, and memory systems. I want to build a local-first personal AI assistant that maintains long-term memory using a knowledge graph. The system should automatically ingest information from emails, meeting notes, documents, and conversations, then use this graph to provide contextual assistance. Design requirements: - Privacy-first: all data stays on-device - Sources: email, calendar, meeting transcripts, documents, voice memos - Graph storage: lightweight, embeddable (no external DB server) - Query latency: < 500ms for context retrieval - LLM integration: works with any local or cloud model Please design: 1. **Knowledge Graph Schema** - Entity types (Person, Organization, Project, Decision, Action Item, Topic) - Relationship types and cardinalities - Temporal modeling - Confidence scoring for extracted facts 2. **Ingestion Pipeline** - Email parsing to entity extraction to graph update - Meeting transcript to key decisions and action items - Document processing to topic extraction and cross-references - Conflict resolution when sources disagree 3. **Retrieval and Context Assembly** - Graph traversal for relevant context given a query - Relevance ranking: recency, frequency, relationship distance - Context window packing strategy - Hybrid retrieval: graph traversal + vector similarity 4. **Memory Lifecycle** - Importance scoring and decay functions - Summarization and compression of old memories - User-controlled forgetting - Graph pruning and maintenance 5. **Implementation Stack** - Recommended libraries and tools - Storage format (SQLite + JSON-LD or alternatives) - Embedding model for hybrid search - Sample code for core operations Provide a practical, implementable design with clear trade-offs discussed.