AuthorsKeer Lu, Liwei Chen, Guoqing Jiang et al.
TL;DR
REAL uses a temporal, confidence-aware memory graph plus reasoning-guided hybrid beam search to boost average long-term memory performance by 22.72% over state-of-the-art baselines.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
LLM memory systems mostly store history as flat text chunks, which cannot track evolving facts or multi-hop relations, leading to shallow retrieval.
REAL highlights that destructive updates and query-agnostic retrieval cause systems to return “I don’t know” and miss implicit evidence, hurting long-term personalization and reasoning.
HOW IT WORKS
REAL builds memory via Atomic Fact Extraction, Incremental Graph Update with Non-Destructive Temporal Evolution, Confidence Stratification and Upgrade, and Exploration Intent Enrichment into a directed property graph.
Think of REAL as a long-term “disk” of structured facts with timestamps and intents, feeding a reasoning engine that selectively loads relevant subgraphs into the LLM’s “RAM” context window.
This design lets REAL preserve parallel fact versions, filter by confidence and time, and perform query-aware hybrid beam search with counterfactual inference, which plain context windows or flat vector stores cannot support.
DIAGRAM
This diagram shows how REAL retrieves a compact evidence subgraph using root entity anchoring, intent decoupling, hybrid beam search, and counterfactual inference.
DIAGRAM
This diagram shows how REAL is evaluated on multiple benchmarks and compared against flat-text and graph-based memory baselines.
PROCESS
Atomic Fact Extraction
REAL segments the conversation stream into turns and uses Atomic Fact Extraction to produce sextuples with entities, relations, valid-time intervals, confidence scores, and exploration intents.
Incremental Graph Update with Non-Destructive Temporal Evolution
REAL inserts each fact into the directed property graph, preserving parallel versions for single-valued and multi-valued relations and closing intervals instead of overwriting when facts change.
Root Entity Anchoring and Intent Decoupling
For a new query, REAL anchors root entities and decouples their exploration intents, producing pairs like (ei, ιi) that guide subsequent traversal in the memory graph.
Semantic Evaluator-Guided Hybrid Beam Search and Counterfactual Inference
REAL runs hybrid beam search with temporal and confidence filters, scores paths by query relevance, logical coherence, and answer sufficiency, and invokes counterfactual inference when expansion becomes unreliable.
KEY CONTRIBUTIONS
Multi-Attribute Memory Construction Mechanism
REAL introduces a temporal, confidence-aware directed property graph with Atomic Fact Extraction and Incremental Graph Update, preserving fact evolution via valid-time intervals and parallel versions.
Reasoning-Enhanced Adaptive Retrieval Strategy
REAL combines Root Entity Anchoring and Intent Decoupling with Semantic Evaluator-Guided Hybrid Beam Search and Counterfactual Inference to discover query-aware memory evidence.
Performance and Effectiveness
REAL achieves an average improvement of 22.72% over flat-text and graph-based memory baselines across dialogue-style and long-form document benchmarks using DeepSeek-V3, Qwen3-32B, and LLaMA-3.3-70B-Instruct.
RESULTS
LoCoMo EM
59.98%
+4.86 over A-MEM
LongMemEval EM
65.92%
+6.29 over Mem0g
PersonaMem EM
37.19%
+4.61 over A-MEM
2WikiMultihopQA EM
43.62%
+3.04 over Vanilla Memory Graph
On LoCoMo, LongMemEval, PersonaMem, HotpotQA, 2WikiMultihopQA, and MuSiQue, REAL consistently surpasses Full Context, flat-text baselines, and graph-based baselines. The 59.98% EM on LoCoMo versus 55.12% EM for A-MEM shows that REAL’s temporal and reasoning-aware memory graph materially improves long-term conversational memory.
BENCHMARK
On LoCoMo, LongMemEval, PersonaMem, HotpotQA, 2WikiMultihopQA, and MuSiQue, REAL consistently surpasses Full Context, flat-text baselines, and graph-based baselines. The 59.98% EM on LoCoMo versus 55.12% EM for A-MEM shows that REAL’s temporal and reasoning-aware memory graph materially improves long-term conversational memory.
BENCHMARK
Exact Match (EM) on LoCoMo for REAL and representative baselines using DeepSeek-V3.
BENCHMARK
Average LLM-as-Judge (LJ) score for REAL and component ablations on Qwen3-32B.
KEY INSIGHT
REAL shows that feeding the entire history is worse than structured retrieval: Full Context on DeepSeek-V3 gets only 35.24% EM on LoCoMo versus 59.98% EM for REAL.
This is surprising because many assume larger context windows always help, but REAL demonstrates that targeted, reasoning-aware memory graphs beat unfiltered long contexts by 24.74 percentage points.
WHY IT MATTERS
REAL unlocks long-term agents that can track evolving preferences, temporal successions, and implicit relations, while repairing missing evidence via counterfactual inference.
Builders can now implement memory systems where facts never silently disappear, queries drive adaptive graph traversal, and LLMs reason over months of interaction without drowning in irrelevant tokens.
Guilin Zhang, Wei Jiang et al.
· 2026
A-MAC scores candidate memories using Utility, Confidence, Novelty, Recency, and Type Prior combined by a learned linear admission policy with Algorithm 1 A-MAC Memory Admission. On the LoCoMo benchmark, A-MAC achieves F1 0.583 and 2644 ms latency, improving F1 by 0.042 and reducing latency by 1187 ms compared to A-mem.
Robbyant Team, Zelin Gao et al.
arXiv 2026 · 2026
LingBot-World combines a Data Engine, Fundamental World Model, Action-Conditioned World Model, and Post-Training causal adaptation to turn a 28B-parameter video generator into a real-time interactive world simulator. On the VBench benchmark, LingBot-World achieves a dynamic degree of 0.8857 versus 0.7612 for Yume-1.5, while also improving imaging quality to 0.6683.
Manoj Madushanka Perera, Adnan Mahmood et al.
· 2026
AgenticAI-DialogGen chains ChatPreprocessor, KnowledgeExtractor, TopicAnalyzer, KnowledgeGraphBuilder, PersonaGenerator, DuelingChat Agent, ConversationValidator, ConversationRefiner, QAGeneration, and PostProcessing to turn raw multi-session chats into topic-guided, persona-grounded conversations with explicit short- and long-term memories. On the TGC / KG memory QA benchmark, Mistral-7B fine-tuned within AgenticAI-DialogGen achieves 87.36 F1, compared to GPT-4’s 83.77 F1 in a zero-shot setting on the same task.
Answers use this explainer on Memory Papers.
Checking…