Summary

Parallel research phase covering three interconnected memory architecture tracks. Consolidates #266 (GraphRAG evaluation), #702 (consistency protocols), and #708 (RL training infrastructure).

Research tracks

1. GraphRAG and Temporal KG evaluation (formerly #266)

• Evaluate GraphRAG for structured org knowledge (relationships, decisions, project history)
• Evaluate Temporal Knowledge Graphs for time-aware memory (what was true when?)
• Compare against current Mem0/Qdrant embedding-only approach
• Deliverable: ADR with benchmark results and migration path (if warranted)

2. Multi-agent memory consistency protocols (formerly #702)

• Shared vs distributed memory architectures for multi-agent coordination
• Consistency models: eventual, causal, strong (trade-offs for agent workloads)
• Conflict resolution when agents write overlapping memories
• Deliverable: protocol spec + reference implementation

3. RL training infrastructure for memory consolidation (formerly #708)

• Evaluate RL training options for the memory consolidation pipeline
• Reward signal design (retrieval quality, agent task success correlation)
• Infrastructure requirements (training loop, evaluation harness)
• Deliverable: feasibility assessment + infrastructure selection

Dependencies

• feat: hybrid search (dense + BM25 sparse) for memory retrieval pipeline #694 (hybrid search) -- provides the retrieval baseline these tracks build on
• feat: memory consolidation upgrades (LLM Merge, Search-and-Ask, diversity penalty) #704 (memory consolidation upgrades) -- provides the consolidation pipeline that track 3 targets