知识图谱驱动的个人 AI 助手记忆系统设计

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.