Skip to main content

How Atulya handles memory conflicts

· 3 min read
Anurag Atulya

Reality changes. A CRM agent learns "Acme Corp is a prospect" in January and "Acme Corp signed a contract" in March. Naive memory either drops history or keeps duplicate facts.

Atulya tracks how knowledge evolves, not just the latest string.

Facts vs consolidated knowledge​

Facts are immediate observations from a single retain. Observations are durable knowledge synthesized over time. Consolidation runs in the background after retain.

When a new fact arrives, _find_related_observations uses recall with:

  • Semantic similarity (embeddings)
  • Token budget via consolidation_max_tokens
  • Strict tag matching (tags_match="all_strict") to avoid cross-scope leakage
recall_result = await memory_engine.recall_async(
    bank_id=bank_id,
    query=query,
    max_tokens=config.consolidation_max_tokens,
    fact_type=["observation"],
    tags=tags,
    tags_match="all_strict",
)

LLM conflict analysis​

_consolidate_with_llm compares the new fact to candidate observations in one call. Context includes observation text, proof counts, source memories, and a chronological time series.

Three merge strategies​

Redundant: Same meaning, different wording → refine one observation.

Contradiction: Opposite claims about the same topic → preserve both states with temporal markers. Updated text explains the arc; when unclear, recency wins.

State update: New information replaces old state → explicit transition language ("used to", "now", "changed from X to Y").

Example: evolving B2B relationship​

Facts over six months:

  1. January: interest in enterprise tier
  2. February: CTO integration meeting
  3. April: $50K contract
  4. September: upgrade to $150K after regional expansion

Naive storage: "Acme is on $150K." Consolidated observation: prospect → $50K customer (April) → $150K tier (September) after expansion.

An agent can answer "How did we land Acme?" with the full arc.

Temporal metadata​

occurred_start = LEAST(occurred_start, COALESCE($7, occurred_start))
occurred_end = GREATEST(occurred_end, COALESCE($8, occurred_end))
mentioned_at = GREATEST(mentioned_at, COALESCE($9, mentioned_at))
  • occurred_start: earliest time the state was true
  • occurred_end: latest related event
  • mentioned_at: last reference

History audit trail​

history.append({
    "previous_text": model["text"],
    "changed_at": datetime.now(timezone.utc).isoformat(),
    "reason": reason,
    "source_memory_id": str(memory_id),
})

Agents can explain why knowledge changed and which fact triggered it.

Tag boundaries​

New observations inherit source fact tags. On update, tags union so contributors retain access. Consolidation never crosses strict tag scopes:

existing_tags = set(model.get("tags", []) or [])
source_tags = set(source_fact_tags or [])
merged_tags = list(existing_tags | source_tags)

Durable vs ephemeral​

Consolidation targets lasting knowledge. "User is in Room 105 right now" is ephemeral; "Acme Corp is in Building B" is durable. That cuts false conflict noise.

Why it matters​

Contradictions do not require picking a single winner. Temporal narrative + history + scoped consolidation gives agents explainable, evolving understanding for preferences, relationships, and compliance-sensitive updates.