AuthorsYuanyi Song, Yukai Wang, Xinbei Ma et al.
TL;DR
REALM uses retrieval-driven memory reconsolidation over a heterogeneous cognitive graph to reach 75.97% accuracy on LoCoMo, +7.17 points over MAGMA.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing long term memory systems mainly update memories with new information using predefined data structures and a fixed retrieval pipeline.
When retrieval is treated as a passive endpoint, agents miss feedback driven reorganization, leading to brittle memory topology and weaker future reasoning.
HOW IT WORKS
REALM combines Unified Cognitive Graph, Memory Construction, Strategy Atom Combination, and Local Topology Reconsolidation into a closed loop memory lifecycle.
Think of REALM as a brain like card catalog: retrieval not only looks up entries but rewires the graph, strengthening or weakening connections after each use.
This retrieval driven reconsolidation lets REALM evolve beyond a static context window, forming usage aware structures that cluster evidence for multi hop and temporal reasoning.
DIAGRAM
This diagram shows how REALM retrieves evidence via strategy atoms and then reconsolidates the activated subgraph after each query.
DIAGRAM
This diagram shows how REALM is evaluated across LoCoMo and LongMemEval with GPT-4o-mini as backbone and judge.
PROCESS
Memory Organization
REALM builds a Unified Cognitive Graph where nodes and edges encode entities, events, episodes, facts, and their relations for later retrieval.
Memory Construction
During integration, REALM’s Memory Construction phase decides to add, merge, or skip candidate units and infers relational edges to update the graph.
Memory Retrieval
Using Strategy Atom Combination, REALM composes seed, expansion, and filtering policies to traverse the cognitive graph and collect evidence Vq.
Memory Reconsolidation
In Local Topology Reconsolidation, REALM applies add, strengthen, or weaken decisions on edges in the activated subgraph Gq to evolve future retrieval paths.
KEY CONTRIBUTIONS
Memory Lifecycle Concept
REALM formalizes long term memory as a closed loop lifecycle over a Unified Cognitive Graph, where retrieval triggers structural updates instead of leaving memory static.
REALM Framework
REALM integrates Memory Construction, Strategy Atom Combination, and Local Topology Reconsolidation into one agentic framework for autonomous organization and adaptive retrieval.
Empirical Validation
REALM reaches 75.97% accuracy on LoCoMo and 65.11% on LongMemEval, with reconsolidation adding up to +6.66 points on single session preference questions.
RESULTS
LoCoMo Average
75.97%
+7.17 over MAGMA
LongMemEval Average
65.11%
+1.31 over Zep
LoCoMo Multi Hop
64.54%
+11.74 over Nemori
LongMemEval Knowledge update
88.89%
+14.49 over MAGMA
REALM is evaluated on LoCoMo and LongMemEval_S, which test multi hop, temporal, open domain, and multi session memory use. These results show REALM’s retrieval driven reconsolidation yields large gains over MAGMA and Zep on challenging long term reasoning tasks.
BENCHMARK
REALM is evaluated on LoCoMo and LongMemEval_S, which test multi hop, temporal, open domain, and multi session memory use. These results show REALM’s retrieval driven reconsolidation yields large gains over MAGMA and Zep on challenging long term reasoning tasks.
BENCHMARK
Average accuracy (%) on LoCoMo using GPT-4o-mini as backbone and judge.
BENCHMARK
Average accuracy (%) on LongMemEval_S using GPT-4o-mini as backbone and judge.
KEY INSIGHT
Enabling reconsolidation increases correct answers by up to +6.66 points on single session preference, while evidence discovery rises by only 3.70% and 5.41%.
This is surprising because we expect better performance mainly from retrieving more memories, but REALM shows reorganizing existing evidence paths matters far more.
WHY IT MATTERS
REALM unlocks long term agents whose memory graphs adapt through usage, strengthening successful reasoning paths and weakening misleading ones automatically.
Builders can now design agents that learn their own retrieval topology over time, moving beyond fixed memory structures and static pipelines that ignore retrieval feedback.
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…