Summary

Proposing standard MCP tools for AI agent memory export/import to enable memory portability across different AI systems. This addresses GDPR Article 20 (data portability) requirements and user data ownership.

Motivation

As AI agents accumulate memory about users (preferences, decisions, context), this data becomes valuable personal information that:

1. Users should own - Not locked to a single provider
2. Should be portable - Move memory when switching AI tools
3. Should be auditable - Inspect what an AI "knows"
4. Complies with GDPR Art. 20 - Right to data portability

Currently, no standard exists for AI memory portability:

• OpenAI's memory is not exportable
• Claude has no persistent memory API
• Memory providers (Mem0, Zep) use incompatible formats

Proposed Tools

1. export_memory

{
  name: "export_memory",
  description: "Export all user memories to portable format",
  inputSchema: {
    type: "object",
    properties: {
      format: { enum: ["json", "yaml", "jsonl"] },
      include_embeddings: { type: "boolean" },
      include_graph: { type: "boolean" },
      since: { type: "string", format: "date-time" },
      redact_pii: { type: "boolean" }
    }
  }
}

2. import_memory

{
  name: "import_memory",
  description: "Import memories from portable format",
  inputSchema: {
    type: "object",
    properties: {
      data: { type: "string" },
      merge_strategy: { enum: ["skip_duplicates", "overwrite", "rename"] }
    }
  }
}

Reference Implementation

We've implemented this in shodh-memory with a proposed open format called MIF (Memory Interchange Format):

• Specification: https://github.com/varun29ankuS/shodh-memory/blob/main/seps/SEP-002-memory-interchange-format.md
• JSON Schema: https://github.com/varun29ankuS/shodh-memory/blob/main/seps/mif-v1.schema.json
• Example: https://github.com/varun29ankuS/shodh-memory/blob/main/seps/example.mif.json

MIF Features

• Memory objects with content, type, importance, entities, embeddings
• Todo/task system with GTD workflow
• Knowledge graph with Hebbian edge weights
• PII detection and redaction
• Optional AES-256-GCM encryption

Benefits for MCP Ecosystem

1. Standardization - Common interface for all memory servers
2. User trust - Data ownership builds confidence in AI tools
3. Interoperability - Enable memory sharing between MCP servers
4. Compliance - Meet regulatory data portability requirements

Discussion

We'd love community feedback on:

1. Should this be a recommended tool interface for memory servers?
2. Should MCP define a standard memory interchange format?
3. What additional considerations for security/privacy?

Related: varun29ankuS/shodh-memory#4