Memory Interchange Format (MIF)
Your AI agent has 6 months of memories in System A. You want to try System B. Without MIF, you lose everything. With MIF:
pip install mif-tools
mif convert mem0_export.json --to shodh -o memories.mif.json
Done. Your memories are portable.
What is MIF?
A vendor-neutral JSON envelope for AI agent memories. Like vCard for contacts or iCalendar for events — a minimal schema so memories move between providers without data loss.
3 required fields. That's it.
{
"mif_version": "2.0",
"memories": [
{
"id": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
"content": "User prefers dark mode across all applications",
"created_at": "2026-01-15T10:30:00Z"
}
]
}
Everything else — memory types, tags, entities, embeddings, knowledge graph, vendor extensions — is optional. Add what you have, ignore what you don't.
Install
# Python
pip install mif-tools # core (zero dependencies)
pip install mif-tools[validate] # with JSON Schema validation
pip install mif-tools[mcp] # with MCP server
# Node.js / TypeScript
npm install @varunshodh/mif-tools
Convert Between Formats
# mem0 → MIF
mif convert mem0_export.json --from mem0 -o memories.mif.json
# MIF → Markdown (Obsidian/Letta style)
mif convert memories.mif.json --to markdown -o memories.md
# Auto-detect source format
mif convert any_memory_file.json -o output.mif.json
# Inspect any memory file
mif inspect memories.json
# Validate MIF document
mif validate memories.mif.json
Python API
from mif import load, dump, convert, MifDocument, Memory
# Load from any format (auto-detects mem0, markdown, generic JSON, MIF)
doc = load(open("mem0_export.json").read())
print(f"{len(doc.memories)} memories loaded")
# Convert between formats in one line
markdown = convert(data, from_format="mem0", to_format="markdown")
# Create memories from scratch
doc = MifDocument(memories=[
Memory(
id="123e4567-e89b-12d3-a456-426614174000",
content="User prefers dark mode",
created_at="2026-01-15T10:30:00Z",
memory_type="observation",
tags=["preferences", "ui"],
)
])
print(dump(doc)) # MIF v2 JSON
# Deep validation (UUIDs, references, timestamps, embedding dimensions)
from mif import validate_deep
ok, warnings = validate_deep(open("export.mif.json").read())
Add MIF to Your MCP Server (10 lines)
from mif import load, dump
# Export handler
def export_memories(user_id: str) -> str:
memories = my_storage.get_all(user_id)
return dump(memories)
# Import handler — auto-detects mem0, markdown, generic JSON, MIF
def import_memories(data: str) -> dict:
doc = load(data)
for mem in doc.memories:
my_storage.save(mem.id, mem.content, mem.created_at)
return {"memories_imported": len(doc.memories)}
Supported Formats
| Format | ID | Auto-detect | Description |
|---|---|---|---|
| MIF v2 | shodh | "mif_version" in JSON | Native format, lossless round-trip |
| mem0 | mem0 | JSON array with "memory" field | mem0 memory exports |
| CrewAI | crewai | JSON array with "task_description" | CrewAI LTMSQLiteStorage exports |
| LangChain | langchain | JSON array with "namespace" + "value" | LangChain/LangMem Item format |
| Generic JSON | generic | JSON array with "content" field | Any JSON memory array |
| Markdown | markdown | Starts with --- | YAML frontmatter (Letta/Obsidian style) |
Full Spec
MIF supports optional fields for rich memory data:
- Memory types —
observation,decision,learning,error,context,conversation, and custom types - Entity references — named entities with type and confidence
- Embeddings — model name, dimensions, vector (reuse or regenerate)
- Knowledge graph — entities and relationships with confidence scores
- Vendor extensions — system-specific metadata preserved on round-trip
- Privacy — PII detection and redaction markers
Full specification: spec/mif-v2.md | JSON Schema: schema/mif-v2.schema.json
MCP Server
Expose MIF tools to any MCP-compatible AI client:
pip install mif-tools[mcp]
mif mcp
Tools: export_memories, import_memories, validate_memories, inspect_memories, list_formats
Adapters & Implementations
| System | Status | Type |
|---|---|---|
| shodh-memory | Production | Built-in HTTP API (/api/export/mif, /api/import/mif) |
| mif-tools (PyPI) | Production | Python package with CLI + MCP server |
| @varunshodh/mif-tools (npm) | Production | TypeScript/Node.js package with CLI |
| mem0 | Adapter ready | Python + npm |
| CrewAI | Adapter ready | Python + npm |
| LangChain | Adapter ready | Python + npm |
| Generic JSON | Adapter ready | Python + npm |
| Markdown (YAML frontmatter) | Adapter ready | Python + npm |
Design Principles
- Minimal — 3 required fields. Everything else is optional.
- Extensible — Unknown fields and vendor extensions MUST be preserved on round-trip.
- Vendor-neutral — The schema doesn't favor any implementation.
- Forward-compatible — Importers MUST ignore unknown fields.
Contributing
We welcome adapter implementations for any memory system. See CONTRIBUTING.md.
Related
- Documentation — Full docs site
- MCP SEP #2342 — Original proposal to the Model Context Protocol
- tower-mcp #531 — Tracking issue in tower-mcp
- shodh-memory — Reference implementation (Rust)
- mif-tools on PyPI — Python package
- @varunshodh/mif-tools on npm — npm package
License
Apache 2.0