Temporal Reasoning

Temporal Reasoning

The graph stores when a fact was true (event-time) separately from when you learned it (ingest-time). Most "what was true on date X?" questions need the former; most audit questions need the latter. Confusing them is the most common subtle wrong answer in assistant memory.

When to Use

• The user asks about a window: "during 2024," "last quarter," "before we started Project Atlas."
• A fact has changed and the user is asking about the old version: "where did Sara work before?"
• Reconstructing a sequence of decisions or moves.
• Reconciling a contradiction by figuring out which fact came first in the world (not just first in the database).

Workflow

1. Reach for the right axis. Window queries → event-time. "When did I find out?" → ingest-time. They are different columns; using the wrong one is silently wrong.
2. Treat event_end = 0 as "still true," not "ended at 0." The query layer expects this convention.
3. For half-open windows, the cutover instant belongs to the next window. A relation that ended at exactly 2025-01-01 00:00 does not overlap "during 2025."
4. Normalize relative time before querying. "Last summer," "before her wedding," "around the move." Convert to a concrete [from_ts, to_ts] first, then query.
5. Surface uncertainty. If the timestamp is approximate, say so in the response — not "in 2023" if you only know "around 2023."

Anti-patterns

• Using last_seen as a proxy for "when was this true?" — it's an ingest signal, not an event signal.
• Returning the most recent matching row when the question wants the historical row.
• Silently downcasting "since spring 2024" to a single point.

Examples

Example 1: "Where did Alice work during 2024?" → Window [2024-01-01, 2025-01-01), event-time overlap. Returns Acme (closed at 2025-01-01) only if it overlaps strictly before that instant.

Example 2: "When did I tell you about my new job?" → Ingest-time query. Look at first_seen on the relation, not event_start of the job itself.