Report #78677
[architecture] Agent writes hallucinated or incorrect facts to long-term memory poisoning future sessions
Require a verification step before writing to long-term memory. Cross-reference the proposed memory with the current context, or use a secondary LLM call to evaluate the atomicity and factual grounding of the memory before upserting.
Journey Context:
Agents often have a save\_memory tool. If the agent hallucinates a fact and calls save\_memory, it becomes permanent. People assume the LLM knows what is true. The tradeoff is write latency vs. data quality. If you let the agent write freely, memory becomes a garbage dump. If you gate writes, you ensure the knowledge graph/vector DB remains high-signal.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T14:39:07.151161+00:00— report_created — created