sethdford Memory Conflict Resolution
When a new observation contradicts a stored fact, the cheap mistake is to overwrite. The right move is to keep both, with time, and let later evidence settle it.
When to Use
- A new fact has the same subject + relation type as an existing one but a different object.
- A confident user statement contradicts a low-confidence inferred fact (or vice versa).
- An observation could be a correction ("Actually it was 2022, not 2023") or an update ("She left Acme last month").
Workflow
- Classify the contradiction first. Single-valued relations (
works_at,lives_in) typically supersede — only one is currently true. Multi-valued ones (knows,interested_in) branch — both can coexist. - For supersession: close the old row with
event_end = cutover_ts, link the new one viasupersedes_id. Don't delete. Yesterday's truth is still part of the user's history. - For low-confidence vs. high-confidence contradiction: flag the new row instead of writing it live. Quarantine until AutoDream review.
- Don't average. Two contradictory facts averaged into a third lukewarm one is worse than either alone — it's now wrong and uncalibrated.
- Surface to the user when ambiguous. "I had her at Acme but you mentioned Globex earlier — has she moved?" beats silently picking one.
Anti-patterns
- Hard-deleting the prior fact ("they updated, so the old one is wrong") — destroys auditability.
- Treating every disagreement as a correction instead of an update.
- Letting the most recent write always win regardless of source quality.
Examples
Example 1: Two months ago user said Sara works at Acme. Today they mention Globex. → SUPERSEDE. Close works_at(Sara, Acme) at today's timestamp, insert new row with supersedes_id. Both visible in window queries: "where did Sara work in March?" still returns Acme.
Example 2: Web scraper says "Alice lives in Tokyo," user-typed history says "Alice lives in Seattle (high confidence)." → FLAG. Do not write to live graph. AutoDream will revisit with more context.