Bi-Temporal Validity

memv tracks knowledge across two independent timelines. Based on Graphiti.

Two Timelines

Timeline	Fields	Question it answers
Event time	valid_at / invalid_at	When was this fact true in the world?
Transaction time	created_at / expired_at	When did memv learn/record this?

Example

A user says "I work at Anthropic" in January, then says "I just moved to OpenAI" in March.

Statement	valid_at	invalid_at	created_at	expired_at
Works at Anthropic	Jan 1	Mar 1	Jan 1	Mar 1
Works at OpenAI	Mar 1	None	Mar 1	None

• The Anthropic fact has invalid_at=Mar 1 (no longer true) and expired_at=Mar 1 (superseded)
• The OpenAI fact has invalid_at=None (still true) and expired_at=None (current record)

Contradiction Handling

When predict-calibrate detects a contradiction, it doesn't delete the old fact. Instead:

1. The old SemanticKnowledge entry gets expired_at set to now
2. A new entry is created with the updated information
3. Full history is preserved

Python

# Check if a knowledge entry is still current
knowledge.is_current()  # expired_at is None

# Check if a fact was true at a specific time
knowledge.is_valid_at(datetime(2024, 2, 1))

Point-in-Time Queries

The bi-temporal model enables querying knowledge as it was at any moment:

Python

from datetime import datetime

# What did we know about user's job in January 2024?
result = await memory.retrieve(
    "Where does user work?",
    user_id="user-123",
    at_time=datetime(2024, 1, 1),
)

# Show full history including superseded facts
result = await memory.retrieve(
    "Where does user work?",
    user_id="user-123",
    include_expired=True,
)

Why Not Just Overwrite?

Overwriting loses information. With bi-temporal tracking:

• You can reconstruct what you knew at any point in time
• You can see how a user's situation evolved
• Debugging is easier — you can trace when and why knowledge changed
• Rollback is possible — superseded records still exist