AuthorsQixuan Sun, Yue Que, Bowei He et al.
TL;DR
CreaMem uses scene-aware Life Scene Memories plus dual episodic–trait encoding to reach 66.40% 4o-Judge accuracy on LongMemEval-S, +6.20 points over MemGAS.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing long term memory systems lack scene awareness, so unrelated life scenes share one retrieval space, causing inflated search and cross scene interference.
They also encode each experience from a single perspective, making it hard to retrieve complementary episodic and trait based views needed for multi hop reasoning.
HOW IT WORKS
CreaMem’s core mechanism combines a Meta Memory Manager, Episodic Memory, three Life Scene Memories, a Core Memory, and a Planner LLM with per memory balanced retrieval.
You can think of CreaMem like a card catalog split into Life, Work, and Interest drawers, plus a chronological diary, all coordinated by a librarian Planner LLM.
This design lets CreaMem route experiences by scene, store dual episodic and trait codes, and retrieve balanced views that a plain context window or flat vector store cannot.
DIAGRAM
This diagram shows how CreaMem processes a user question through memory selection, per memory balanced retrieval, and Reciprocal Rank Fusion.
DIAGRAM
This diagram shows how CreaMem is evaluated on LoCoMo and LongMemEval-S and how ablations modify memory components.
PROCESS
Meta Memory Manager
The Meta Memory Manager routes each incoming message into relevant Life, Work, Interest, Episodic, and Core Memory components, following a wide entry strict filtering policy.
Memory Storage
Memory Storage uses component specific LLM extractors to create episodic tuples and scene traits, merging or inserting entries into SQLite backed Episodic Memory and Life Scene Memories.
Memory Retrieval
Memory Retrieval lets the Planner LLM select Core, Episodic, Life, Work, and Interest, then performs per memory hybrid BM25 plus embedding retrieval with balanced sampling.
Reciprocal Rank Fusion
Reciprocal Rank Fusion globally re ranks candidates from all memories, and CreaMem inserts the top N entries into the prompt for response generation.
KEY CONTRIBUTIONS
Scene aware memory partitioning
CreaMem partitions memory into Episodic Memory, three Life Scene Memories for Life, Work, Interest, and a Core Memory, reducing cross scene interference during retrieval.
Dual episodic and trait encoding
CreaMem stores each experience as a structured episodic tuple and a scene specific trait, enabling cross memory synergy via per memory balanced sampling at retrieval time.
Improved multi hop reasoning
On LoCoMo and LongMemEval S, CreaMem improves 4o Judge accuracy to 54.61% and 66.40%, with ablations showing that cross memory synergy accounts for most of the gain.
RESULTS
4o-J
66.40%
+6.20 over MemGAS on LongMemEval-S
F1
20.71%
vs 20.38% for MemGAS on LongMemEval-S
4o-J
54.61%
+8.99 over HippoRAG 2 on LoCoMo
Avg. Tokens
7,250
uses far fewer tokens than Full History at 103,137 on LongMemEval-S
On LoCoMo and LongMemEval S, which test long term conversational memory and multi session retention, CreaMem consistently achieves the highest 4o Judge accuracy among all baselines. These results show that scene aware dual coded memory organization yields better retrieval than flat or purely structural systems at similar or lower token budgets.
BENCHMARK
On LoCoMo and LongMemEval S, which test long term conversational memory and multi session retention, CreaMem consistently achieves the highest 4o Judge accuracy among all baselines. These results show that scene aware dual coded memory organization yields better retrieval than flat or purely structural systems at similar or lower token budgets.
BENCHMARK
4o-Judge Acc (%) on LongMemEval-S for CreaMem and strong baselines.
BENCHMARK
4o-Judge Acc (%) on LoCoMo for CreaMem and strong baselines.
KEY INSIGHT
CreaMem’s Full System beats Episodic only by 9.2 points on overall Gemini 3 Judge accuracy, while actually using fewer tokens per query.
This is surprising because many assume that more raw episodic context always helps, but CreaMem shows that structured dual encoding and balanced sampling can be both shorter and more accurate.
WHY IT MATTERS
CreaMem unlocks long term agents that can reason over months of interactions using scene aware, dual coded memories instead of monolithic history dumps.
Builders can now design agents that keep separate Life, Work, and Interest personas, answer temporal and trait questions jointly, and scale memory without exploding context windows.
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…