This paper introduces DISCERN (DISambiguation and Cross-session Episodic memoRy Navigation), a framework that reimagines how LLM agents handle long-term, persistent memory. Rather than treating memory as a retrieval problem, the authors argue that the real challenge lies in resolution: correctly identifying entities, handling superseded facts, and knowing when to stay silent.
TL;DR
DISCERN moves agent memory from "retrieval-centric" to "resolution-centric" by following a simple rule: preserve at write, resolve at read. By maintaining a strictly additive, bi-temporal provenance graph and using an agentic navigation policy, it significantly outperforms existing systems in entity disambiguation and calibrated abstention.
What Problem Does This Paper Address?
In enterprise settings, agents accumulate memory over dozens of sessions. This leads to three major issues:
- Supersession: Facts change over time (e.g., a deal stage moves from "Negotiation" to "Closed Won").
- Entity Collision: Similarly named entities (e.g., "Zephyron Cloud Solutions" vs. "Zephyron Cloud Systems") blur together in vector space.
- The "Commitment" Problem: Standard RAG systems often hallucinate or blend facts because they lack a formal mechanism to decide if they have enough evidence to answer, or if they should clarify or abstain.
The authors demonstrate that for modern LLMs, retrieval (finding the right information) is largely solved, with baselines hitting 93-98% recall. The bottleneck is resolution—interpreting that retrieved data correctly.
Methods
DISCERN splits the memory process into two distinct paths:
1. Preserve at Write (The Store)
Instead of merging similar entities or overwriting old facts, DISCERN builds a strictly additive bi-temporal provenance graph.
- Never Merge: Confusable entities are linked via
SIMILAR_NAMEedges but kept as distinct nodes. - Soft Invalidation: When a value changes, the old value is marked as "superseded" but retained. This allows the agent to answer questions about historical states (e.g., "What was the revenue before the acquisition?").
2. Resolve at Read (The Policy)
At query time, DISCERN doesn't just do a similarity search. It uses an agentic loop to navigate the graph using a suite of deterministic, zero-LLM tools (like find_similar_entities and check_temporal).
The process culminates in one of four governed terminal actions:
- ANSWER: A confident, cited response.
- UPDATE: Returning the latest value in a chain of changes.
- CLARIFY: Triggered when the agent finds "confusable twins" and needs more detail.
- ABSTAIN: A verified refusal when information is missing.
Datasets
The authors evaluate DISCERN on two primary benchmarks:
- ENT-310 (New): A synthetic enterprise sales-agent benchmark featuring 310 queries designed to test retention, currency, contamination (leaking facts between similar entities), and entity resolution. It includes 100 unanswerable "trap" questions to test abstention.
- LoCoMo: An existing public benchmark of long-term personal conversations, used here to test out-of-domain generalization.
Key Insights and Ideas
- Resolution > Retrieval: The "aha moment" is that even when a RAG system retrieves the correct passage, it often fails to resolve which specific entity or time-state the user is asking about.
- Lossless Writing is Safer: Write-time merging is irreversible. By deferring the decision to merge or overwrite until read-time, the system remains flexible and avoids permanent data corruption.
- Graph-Decidable Commitment: By using a labeled graph, the decision to answer or abstain becomes a decidable function of the graph structure (e.g., "If I find two viable candidates, I must CLARIFY") rather than a fuzzy LLM judgment.
Key Results
- Entity Resolution: On ENT-310, DISCERN achieved a score of 0.800, while no other baseline (including GraphRAG and HippoRAG2) exceeded 0.40.
- Calibrated Abstention: DISCERN is the best-calibrated system for knowing when not to answer, achieving an F1 of 0.924 on unanswerable questions (compared to 0.802 for the best baseline).
- Generalization: On the LoCoMo dataset, DISCERN led in macro answer correctness (0.663 vs 0.638) without any domain-specific tuning.
Limitations and Future Work
- Latency: The agentic navigation loop is thorough but slow, averaging 4.1 turns and 5.7 tool calls per query (mean latency of 7.6s).
- Backbone Sensitivity: The system relies heavily on the LLM's ability to follow complex tool-use instructions. Performance dropped significantly when using weaker open-weight models compared to Claude 3.5 Sonnet.
- Future Work: The authors suggest training a specialized, smaller model to handle the navigation policy to reduce latency and cost.
Summary updated