Contradictions and updating facts

The memory contains «prefers code examples» from March. In October the user says «I would rather have the formulas first». Now there are two contradictory facts.

Detection: at consolidation, new facts are compared with the existing ones within the same type and subject. An NLI model or an LLM decides whether they are contradictory.

The resolution, in priority order:

1. Explicit over derived. Something the user has expressly said beats something the system guessed from behaviour.
2. Newer over older — but only when both are equally explicit.
3. More support over less. Three episodes beat one.
4. A draw → ask the user, or keep both with a low confidence and avoid acting on them.

Rule four is the most important: guessing wrong in a contradiction is worse than asking.

from datetime import datetime

SOURCE_WEIGHT = {"explicit": 3.0, "confirmed": 2.0, "derived": 1.0}

def resolve_conflict(a: dict, b: dict, now: datetime) -> dict:
    """Returns {'keep': fact, 'remove': fact} or {'ask': (a, b)}."""
    def score(f):
        age_d = (now - datetime.fromisoformat(f["ts"])).days
        freshness = 0.5 ** (age_d / 180)
        return SOURCE_WEIGHT[f["source"]] * freshness * (1 + 0.2 * len(f["sources"]))
    pa, pb = score(a), score(b)
    if abs(pa - pb) < 0.25 * max(pa, pb):        # too close → do not guess
        return {"ask": (a, b)}
    winner, loser = (a, b) if pa > pb else (b, a)
    return {"keep": winner, "remove": loser, "motivation": f"{score(winner):.2f} > {score(loser):.2f}"}

# Always log which fact replaced which — otherwise it is impossible to debug
# when the agent suddenly behaves differently towards a user.

The trap to avoid: silent deletion. When a fact is replaced it should be visible in the log — and preferably to the user in the settings. A memory that changes invisibly is impossible to trust.