Memory Landscape

The Agent Memory market has formed several patterns: lightweight fact memory, temporal graphs, state checkpoints, multi-agent shared memory, and full RAG platforms.

Product patterns

Pattern	Representative products	Strength	Tradeoff
Lightweight facts	Mem0	Simple SDK, low intrusion, cross-session retrieval	Complex temporal conflicts require extra engineering
Temporal graph	Zep	Strong at state evolution, relationships, and context assembly	Heavier model and higher abstraction cost
Memory OS	Letta	Agents can self-edit long-term memory	Highly invasive control flow
State checkpoint	LangGraph	Snapshots, rollback, and fault tolerance	Semantic retrieval needs external systems
Multi-agent memory	CrewAI	Easy shared context across roles	Black-box behavior and token growth
Full RAG platform	Graphlit	Multimodal ETL and retrieval pipeline	Heavy architecture and higher cost for lightweight memory

Shared market gaps

Existing memory products often share these gaps:

• Memory write and recall are not transparent enough.
• Developers cannot easily explain which memory shaped an answer.
• Privacy, erasure, and lifecycle governance are not always first-class.
• Memory is not deeply integrated with identity and web action.

GUM direction

GUM's differentiation is not being another memory store. It places memory inside the SAK ecosystem:

• GenAuth provides identity-based behavior memory.
• Web Agent provides search and action context.
• Provenance records provide memory attribution and explainability.

White-box memory

GUM should emphasize white-box memory:

• Which memory was recalled.
• Why it was recalled.
• Which conversation, search, or action produced it.
• Whether the user confirmed, changed, or deleted it.
• How it influenced the final answer.