The MCP spec currently focuses on tools, resources, and prompts. As more projects build agent memory systems, I think there's a case for standardizing a memory capability type within MCP.

Use cases:

• Cross-session agent context persistence
• Shared memory between multiple agents in a team
• Structured memory graphs (knowledge graphs, entity relationships)
• Memory retrieval with semantic search

There are already projects building this (e.g., Awareness provides a memory cloud for AI agents with MCP support), but without a standard capability type, each implementation uses ad-hoc patterns.

Would the MCP spec team consider adding a memory capability alongside tools/resources/prompts? I think it would help the ecosystem converge on interoperable memory patterns.

Could start with a lightweight proposal defining:

• Memory read/write primitives
• Namespace/scoping rules (per-agent, per-session, shared)
• TTL and eviction semantics

Happy to help draft if there's interest.