On ENT-310, a synthetic enterprise sales-agent memory benchmark of 310 queries, Retrieval-Augmented Generation systems saturate retention and currency scores between roughly 0.93 and 0.98, yet their entity resolution accuracy collapses below 0.40. Standard enterprise architectures assume that long-term agent interactions fail because the model cannot pull the right context out of vector stores. The operational reality recorded by Amazon researchers Chirag Ahuja, Vladislav Tanov, Ramcharan Kakarla, and Jonathan Taws shows that recall is essentially solved, while corporate deployments bleed accuracy because agents cannot disambiguate similarly named entities or determine when to decline unanswerable questions.
Eliminating memory hallucinations across dozens of customer sessions without running up massive inference bills remains the central operational bottleneck for autonomous systems. The proposed answer separates architecture across time: abandoning standard RAG pipelines for bi-temporal graphs fixes conflicting corporate records while eliminating unnecessary large language model calls. Amazon frames this architectural pattern around DISCERN, an agent-memory system operating under an explicit rule to preserve at write and resolve at read.
The Retrieval Trap and Entity Collisions
When a sales representative asks where a deal landed three sessions prior, an enterprise agent faces a warm-start task where the answer lives in earlier sessions. Because a single context window cannot hold months of advisor interactions across dozens of client accounts, traditional architectures rely on persistent storage split into distinct write and read paths. Standard write paths extract entity-attribute-value facts, such as recording fictitious revenue of $5.1B for Zephyron Cloud Solutions, alongside rolling session summaries under per-actor namespaces. At read time, a retriever pulls the top 10 most similar records from both stores, while an optional query-enrichment model call attempts to pinpoint the entity.
"Retention and currency are near-saturated, roughly 0.93 to 0.98 for every competent system, so recall is essentially solved. What remains unsolved is resolution: which entity a query refers to, which time-state is current, and whether to commit an answer at all or decline when evidence is insufficient."
This workflow conceals severe architectural failure modes. Standard baseline systems rely on one-shot disambiguation through a single post-retrieval call without iterative verification loops. Crucially, these systems lack a terminal-state contract: they always return an answer even when the prompt is unanswerable from stored memory. Standard pipelines do not distinguish between ANSWER, UPDATE, CLARIFY, and ABSTAIN states, allowing stale, hallucinated, and blended data to contaminate outputs while remaining completely invisible to conventional retrieval metrics. Simple query enrichment and vector expansion cannot prevent these collisions.
The benchmark numbers prove that vector indexing alone cannot bridge the gap. In fact, standard RAG readers paired with semantic and summary stores match advanced memory setups on pure retrieval, but they fail when names collide in embedding space or facts contradict earlier records. Resolving time-states and ambiguous entities requires structured organization before queries arrive, rather than forcing an expensive model to improvise clarity during execution.
Structured Provenance Over Token Ingestion
Amazon implements DISCERN by strictly dividing labor across the lifecycle. At write time, the framework constructs an immutable per-actor event log as a single source of truth and consolidates observations into a derived bi-temporal provenance graph. This write path is additive: it never merges confusable entities and never discards superseded values. At read time, an agentic loop executes over a zero-LLM tool surface, navigating the graph and committing to exactly one of four gated terminal actions: ANSWER, UPDATE, CLARIFY, or ABSTAIN.
"On ENT-310, DISCERN is the best-calibrated system on abstention by a wide margin (F1 0.924 versus at most 0.802 across eight baselines) and leads macro answer correctness under equal pillar weighting (0.861 versus 0.806 for the strongest baseline), a lead driven by entity resolution (0.800, where no baseline exceeds 0.40)."
Evaluations conducted across a four-judge LLM-as-a-Judge council reveal the architectural payoff. On ENT-310, which contains 210 answerable and 100 unanswerable queries, DISCERN achieved an abstention F1 score of 0.924, compared to a maximum of 0.802 across eight baseline systems. Under equal pillar weighting, DISCERN reached macro answer correctness of 0.861 versus 0.806 for the strongest baseline. That outperformance stems from entity resolution, where DISCERN scored 0.800 while no baseline exceeded 0.40. When transferred without tuning to the LoCoMo benchmark corpus, the identical pipeline led macro answer correctness at 0.663 versus 0.638 and maintained calibration on false-premise abstentions.
Persistent enterprise memory systems cannot scale on brute vector similarity when operational data continuously changes and overlaps. Moving logic from expensive inference-time calls into deterministic bi-temporal graphs provides a viable path to prevent hallucinated commitments in autonomous agents. Infrastructure architectures that fail to formalize state contracts will face mounting validation costs as multi-session histories expand.
What it signals
Vector storage fetches the right records, but an LLM-agent easily mixes up lookalike client profiles and fabricates an answer instead of admitting it lacks context (standard pipelines simply never learned when to abstain). While the industry weighed the risk of memory and RAG poisoning back in mid-May, the Amazon team proved that autonomous setups can derail entirely on their own conflicting data. Will the DISCERN framework survive beyond the 310 benchmark queries without ballooning LLM inference costs during read operations?
Our stories: AI Agent Safety: The Hidden Dangers of Long-Term Memory (May 19).