Source-linked AI summary
CreaMem: A Scene-Aware Memory Architecture for Personalized Agents
Qixuan Sun, Yue Que, Bowei He, Jin Guo, Dihang Yang, Wenchang Situ, Chen Ma
TL;DR
Long-term memory systems lack scene awareness and often encode experiences from only one perspective, limiting organization and complementary retrieval. CreaMem partitions memory by life scenes, dual-codes experiences episodically and as scene-specific traits, and balances retrieval across memories. On two long-term memory benchmarks, it improves QA accuracy across metrics, with especially strong multi-hop reasoning gains.
Problem
Existing memory systems lack life-scene organization and rarely encode each experience from both episodic and trait-based perspectives.
Method
CreaMem partitions memory into Life Scene Memories, dual-codes experiences in episodic and trait forms, and uses per-memory balanced sampling during retrieval.
Results
CreaMem improves overall QA accuracy and multi-hop reasoning on two long-term memory benchmarks, with gains consistent across evaluation metrics.
Takeaways & Limitations
The results suggest that organizing memory, rather than merely supplying more context, is a key lever for coherent long-term agents.
Takeaways & Limitations
The fixed Life/Work/Interest taxonomy is deliberately coarse and may not transfer uniformly across users, cultures, or specialized domains.
Abstract
from arXiv · showhide
Long-term memory is a core capability for personalized LLM agents. To support it, existing memory systems organize information using various criteria such as topic segments or summary hierarchies. However, we identify two major limitations in these designs. First, they lack scene awareness: memories from unrelated life scenes share the same retrieval space, which inflates the search space and introduces cross-scene interference. Second, they encode each memory from a single perspective, making it difficult to retrieve complementary views of the same event. In this paper, we propose the CreaMem architecture, which enables scene-aware memory organization by partitioning memory into several Life Scene Memories to reduce cross-scene interference at retrieval. To go beyond the single perspective and achieve cross-memory synergy, entries are dual-coded from both episodic and trait-based perspectives within each memory. We further devise a permemory balanced sampling strategy at retrieval time. Extensive experiments on two long-term memory benchmarks show that CreaMem improves QA accuracy across all evaluation metrics, with particularly large gains on multi-hop reasoning performance, validating scene-aware partitioning and cross-memory synergy. To enhance reproducibility, we release our code in a public GitHub repository.
1 Introduction
CreaMem addresses two gaps in long-term memory organization: the lack of user-life-scene awareness and the single-perspective encoding of experiences. It partitions memories by scene, dual-codes experiences episodically and as traits, and combines these views during retrieval.
- Long-term memory organization and retrieval are central to sustaining coherent behavior across extended interactions.
- Existing systems use flat storage, explicit structures, or memory partitioning, but do not organize memories by users’ life scenes.
- Each experience is stored as both a timeline entry in Episodic Memory and a trait entry in the relevant scene memory.
- A Meta Memory Manager routes incoming messages to relevant scenes, while per-memory balanced sampling combines timeline and scene-specific views in returned context.
- CreaMem partitions memory into Life, Work, and Interest scenes alongside a dedicated Episodic Memory, reducing cross-scene interference at retrieval.
- Experiments on two long-term memory benchmarks show improved overall QA accuracy and multi-hop reasoning, with ablations attributing most gains to cross-memory synergy.
2 Related Work
Prior work organizes agent memory through context windows, undifferentiated retrieval pools, explicit structures, or specialized components. CreaMem instead draws on cognitive theories that emphasize lifetime contexts and complementary episodic-semantic perspectives.
- LLM agents struggle to maintain long-term coherence when fixed-length contexts cannot preserve continuity across temporally separated dialogues.
- Flat-storage systems retrieve from undifferentiated pools where unrelated entries compete for retrieval slots.
- Memory-partitioning architectures decompose memory along axes such as temporal scale or cognitive function.
- Cognitive foundations describe autobiographical memories as organized by lifetime contexts rather than abstract categories.
- Episodic and semantic memory encode the same experience from complementary perspectives and cooperate during retrieval.
- CreaMem applies these principles by storing each experience as a timeline entry in Episodic Memory and a trait entry in its relevant scene memory.
3 CreaMem
CreaMem organizes long-term agent memory into scene-specific and episodic components, with storage and retrieval mechanisms that preserve complementary temporal and trait perspectives. Its retrieval pipeline selects relevant memories, balances per-memory candidates, and fuses them globally.
- Memory Components: CreaMem comprises Episodic, Life, Work, Interest, and Core Memories, separating time-anchored events, scene-specific traits, and holistic profile information.Episodic Memory stores concrete past events, while Life Scene Memories store traits within personal, professional, academic, and leisure contexts.
- Memory Storage: The Meta Memory Manager routes each incoming message to relevant scene memories and adds a paired timeline entry to Episodic Memory.Because messages may span multiple scenes, the same message can be dispatched to several memories simultaneously.
- Memory Storage: Life Scene entries represent distilled, scene-conditional traits extracted from one or more dialogue messages and stored using a shared schema.Each trait includes a unique identifier, extracted content, an importance score, and an embedding.
- Memory Storage: Episodic entries preserve timestamps, actors, event types, summaries, full context, and dense embeddings for grounding retrieval in specific past moments.A single message may yield multiple structured episodic entries.
- Memory Retrieval: For each question, a Planner LLM selects relevant memories and keywords, after which BM25 and embedding retrieval run in parallel within each selected memory.Each retrieval path returns top-k candidates from its memory.
- Memory Retrieval: Balanced sampling gives every selected memory the same number of candidates before deduplication and global Reciprocal Rank Fusion ranking.The top-N fused candidates enter the response prompt, with selected memories re-queried when the context is judged insufficient.
4 Experiments
CreaMem is evaluated against representative long-term memory baselines on LoCoMo and LongMemEval-S using multiple accuracy and similarity metrics. It achieves the strongest overall results, with ablations showing benefits from scene partitioning and dual encoding, while retrieval performance depends on budget and can fail on individual questions.
- Main Results: Structure-imposing methods remain limited when unrelated life scenes share one retrieval pool, producing cross-scene interference.Temporal-scale partitioning likewise leaves unrelated scenes co-occurring within the same tier.
- Main Results: CreaMem achieves the best performance across both datasets and all reported metrics, surpassing the strongest baselines on each benchmark.The lead is consistent across F1, BLEU-4, ROUGE-1/2/L, and BertScore.
- Main Results: CreaMem’s gains arise from memory organization and retrieval rather than a larger context budget.Its token consumption is comparable to strong structured baselines on LoCoMo and below most baselines on LongMemEval-S.
- Ablation Study: 3-scene partitioning outperforms 1-scene storage by 2.3 points under identical content and retrieval budgets.The tested 3-scene Life/Work/Interest configuration achieves the best balance, whereas finer partitions over-fragment memory.
- Retrieval Sensitivity: Retrieval accuracy is non-monotone in k: 4o-Judge peaks at k=30, then declines at k=40 as token cost grows nearly linearly.Beyond a modest budget, additional entries can act as noise rather than signal.
- Failure Analysis: CreaMem still fails on some questions because linked entries may be incompletely retrieved or retrieved evidence may not be combined.The failure analysis motivates links to underlying dialogue turns, fallback to relevant segments, and explicit evidence combination.
5 Conclusion
CreaMem is a scene-aware memory architecture that combines life-scene organization with episodic and trait-based views of experience. Experiments show gains in accuracy and multi-hop reasoning at comparable token cost.
- CreaMem partitions memory along user life scenes alongside Episodic and Core Memory.
- Each experience is dual-encoded as a timeline entry and a trait entry, with views combined through per-memory balanced sampling.
- Experiments on LoCoMo and LongMemEval-S show consistent gains in accuracy and multi-hop reasoning at comparable token cost.
- The results suggest that memory organization, rather than context quantity, is a key lever for coherent long-term agents.
Limitations
CreaMem’s limitations concern the scope of its fixed scene taxonomy, additional retrieval overhead, and unresolved evaluation challenges for LLM-extraction memory systems.
- The fixed Life/Work/Interest taxonomy is coarse and may not transfer uniformly across users, cultures, or specialized domains.
- End-to-end latency averages 2.9 seconds per query versus 1.98 seconds for A-Mem, while sufficiency checks can add retrieval rounds, latency, and tokens.
- Write-time costs remain difficult to compare because methods differ in batching, prompts, and the abstraction level of stored entries.
- Retrieval metrics can misalign with reference evidence when memory entries summarize or combine multiple dialogue turns.
- Extraction behavior depends on prompt and model choice, motivating standardized write-time accounting and method-agnostic extraction protocols.
A Supplementary Analyses
Supplementary analyses show that CreaMem leads several LoCoMo categories, transfers across backbones, and gains primarily from scene-aware dual encoding rather than allocation or fusion alone.
- CreaMem leads Overall, Multi-hop, and Temporal accuracy under the matched GPT-4o-mini/GPT-4o protocol, while SeCom and MemoryOS lead other categories.
- CreaMem’s margins over the strongest baseline are 2.83 points on Multi-hop and 19.93 points on Temporal questions.
- With Qwen3.6-35B-A3B, CreaMem leads by 5.97 points overall and remains strongest on Multi-hop, Open-domain, and Temporal questions.
- Full CreaMem reaches 53.24%, versus 44.03% for Episodic-only and 40.32–41.43% for scene-only variants.
- Dynamic allocation reaches 55.26% versus 55.00% for balanced sampling, adding 0.69 LLM calls per question.
- Representative failures distinguish incomplete retrieval of linked evidence from failure to combine evidence already present in context.
B Prompt Templates
The supplementary prompt templates specify CreaMem’s routing, retrieval, extraction, QA, and judging procedures, alongside the evaluation and audit materials used to assess them.
- Storage routing: The Meta Memory Manager routes batched user messages to one or more memory types and then finishes the memory update.
- Retrieval planning: The retrieval planner outputs keywords and selected memory components, including episodic memory for events and scene memories for traits or preferences.
- Retrieval planning: On parse failure or empty output, retrieval falls back to all four components using the question itself as the keyword.
- Evaluation materials: Table 5 presents representative questions that CreaMem answered incorrectly while a baseline answered correctly, distinguishing failure stages rather than prevalence.
- Scene extraction: Scene Memory extractors cover Life, Work, and Interest, with scene-specific definitions and entries represented as content paired with an importance score.
- Scene extraction: Extraction rules create separate entries for distinct facts, preserve proper nouns and concrete details, and ignore content outside the assigned scene.