RFC: Memory Interchange Format (MIF) - Portable AI Memory Standard

[varun29ankuS]

📋 Overview

We're proposing MIF (Memory Interchange Format) - an open standard for portable AI agent memory. As AI agents become persistent collaborators, their memory becomes valuable personal data that users should own and control.

🎯 Goals

1. Portability - Move memory between AI systems (Claude → GPT → local agents)
2. Backup/Restore - Full-fidelity memory preservation
3. Interoperability - Share memory across agents (with consent)
4. Auditability - Inspect what an AI "knows" about you
5. Privacy - GDPR Article 20 compliance (data portability)

📄 Specification

The full specification is in SEP-002: Memory Interchange Format

Key Features

• JSON-based format with optional YAML frontmatter
• Memory objects with content, type, importance, entities, embeddings
• Todo system with GTD workflow support
• Knowledge graph with Hebbian edge weights
• PII redaction for privacy-safe exports
• Encryption (AES-256-GCM) for sensitive exports

Example

{
  "mif_version": "1.0",
  "memories": [{
    "id": "mem_abc123",
    "content": "User prefers Rust for systems programming",
    "type": "Learning",
    "importance": 0.85,
    "entities": [{"text": "Rust", "type": "TECHNOLOGY"}]
  }]
}

🔄 Implementation Status

• ✅ Export endpoint (/api/export)
• ✅ Import endpoint (/api/import)
• ✅ Full graph export with Hebbian edges
• ✅ PII detection and redaction
• ⬜ JSONL streaming for large exports
• ⬜ Cross-system import (Mem0, Zep)
• ⬜ MCP tools for export/import

🤔 Discussion Questions

1. Memory types: Are the 13 defined types sufficient? Should we add more?
2. Embeddings: Should we mandate a specific embedding model for portability?
3. Graph format: Is adjacency list the right choice vs edge list?
4. Encryption: Should encryption be mandatory for exports containing PII?
5. Versioning: How should we handle migration between MIF versions?

🚀 Next Steps

1. Gather community feedback on this RFC
2. Submit as proposal to MCP ecosystem
3. Reach out to other memory providers (Mem0, Zep, etc.)
4. Work toward industry-wide adoption

Your feedback shapes this standard! Please comment with thoughts, concerns, or suggestions.