Summary

The spec describes two research-direction org memory backends beyond the implemented Hybrid Prompt + Retrieval (Backend 1). These are explicitly marked as "not planned implementations; research directions" in the spec, but are documented features behind the OrgMemoryBackend protocol that could be implemented if Backend 1 proves insufficient.

Design Spec Reference

• §7.4 Shared Organizational Memory — Research Directions (M5+)
  • Backend 2: GraphRAG Knowledge Graph
  • Backend 3: Temporal Knowledge Graph

Scope

• GraphRAG: Entities + relationships in a knowledge graph, multi-hop reasoning (e.g. "Litestar is our standard" → linked to → "don't use Flask")
• Temporal KG: Like GraphRAG but tracks how facts change over time

Notes

• These are research directions per the spec — evaluate if/when Hybrid Prompt + Retrieval proves insufficient
• Both implement OrgMemoryBackend protocol (already defined)
• The selected memory layer backend may provide optional graph memory capabilities that reduce implementation effort — evaluate at implementation time

Research Context

Always On Memory Agent (Google, 2026-03-07) — Challenges the need for separate graph DB:

• SQLite + LLM consolidation (three-agent pattern: Ingest/Consolidate/Query) may suffice without Mem0/Zep/Neo4j
• Scheduled 30-min consolidation, LLM semantic reranking
• Validates SQLite-first approach
• Scale concerns: unproven beyond 1k sessions — graph DB may still be needed at scale
• RETHINK: "custom" path (SQLite + LLM + tiered scoring) could be a simpler alternative to Neo4j + Qdrant external

Agents Need Vector Search (VentureBeat/Qdrant, 2026-03-14):

• DeepMind: single-vector embeddings hit hard ceiling on relational complexity
• Multi-vector / hybrid approaches architecturally necessary at scale
• pgraph (PostgreSQL extension) could serve GraphRAG direction without separate Neo4j