CreaMem: A Scene-Aware Memory Architecture for Personalized Agents

• The introduction frames memory organization and retrieval as central to coherent long-term behavior because retaining information alone does not ensure the right information can be retrieved across extended interactions (Packer et al., 2023; Park et al., 2023).Source: §1 · p. 1
• Existing systems use flat storage, explicit structures, or specialized memory components, but none organize by life scenes and few encode experiences from both episodic and trait-based perspectives.Source: §1 · p. 1
• Each experience becomes an episodic timeline entry and a trait entry in its relevant scene, preserving context-dependent preferences and offering complementary evidence for broader queries.A temporal query can use the episodic entry alone, while a broader query can draw on both the event and its scene-specific trait.Source: §1 · p. 2
• Across two long-term memory benchmarks, CreaMem improves overall QA accuracy and multi-hop reasoning, while ablations attribute most gains to cross-memory synergy rather than any single component.Source: §1 · p. 2
• CreaMem partitions memory into Life, Work, and Interest scenes plus Episodic Memory, and uses per-memory balanced sampling to combine timeline and scene-specific views in returned context.A Meta Memory Manager routes each incoming message to the relevant scene memory and adds a paired Episodic Memory entry; retrieval queries all memories.Source: §1 · p. 2

2.1 Memory for LLM Agents

• Fixed-length context struggles with long-term coherence, while full-context methods scale poorly as dialogue histories exceed the context window (Wang et al., 2024; Guo et al., 2024; Wu et al., 2026; Gao et al., 2024; Liu et al., 2024; Paulsen, 2025).Source: §2.1 · p. 2
• Flat-storage and structured memory methods still retrieve from a single pool, where unrelated contexts compete for retrieval slots (Lewis et al., 2020; Wang et al., 2025; Sarthi et al., 2024; Pan et al., 2025).The passage also describes knowledge graphs and note-based links as structures that retain this single-pool limitation.Source: §2.1 · p. 2
• Memory-partitioning architectures organize components by temporal scale or cognitive function, but leave scene interference and single-perspective encoding unaddressed (Kang et al., 2025; Wang and Chen, 2025).Source: §2.1 · p. 2

2.2 Cognitive Foundations

• Conway’s Self-Memory System organizes autobiographical memories by lifetime contexts, while Tulving’s account describes episodic and semantic memory as complementary perspectives that cooperate in retrieval (Conway and Pleydell-Pearce, 2000; Conway, 2005; Tulving, 1972, 2002).Source: §2.2 · p. 3
• CreaMem applies these ideas by storing each experience as a timeline entry in Episodic Memory and a trait entry in the relevant Life Scene Memory.Source: §2.2 · p. 3

3.1 Overview Architecture

• CreaMem comprises episodic, life-scene, and holistic profile memories; incoming messages are routed to scene memories and paired with timeline entries.The Meta Memory Manager writes each experience from temporal and trait perspectives.Source: §3.1 · p. 3

3.2 Memory Components

• Time-anchored event narratives complement scene-conditional traits across personal, work, and leisure domains, alongside a stable cross-scene profile and question-driven component selection.The Meta Memory Manager can route one message to multiple relevant scenes; the Planner LLM selects components for a question.Source: §3.2 · pp. 3–4

3.3 Memory Storage

• Incoming messages are classified and routed broadly across relevant components; LLM extractors propose entries that embedding lookup helps merge with existing entries or insert into SQLite.When scene classification is uncertain, multiple components may be triggered to prevent silent information loss.Source: §3.3 · p. 4

3.4 Memory Retrieval

• For each selected memory, CreaMem applies BM25 and cosine-similarity retrieval, combines per-memory candidates so each memory contributes the same number, then globally reranks the deduplicated pool with Reciprocal Rank Fusion.The Planner LLM selects relevant memories and a set of 3 to 6 retrieval keywords.Source: §3.4 · pp. 4–5

4.1 Evaluation Setup

• Evaluation uses LoCoMo’s 1,540 questions and LongMemEval-S’s 500 questions, with GPT-4o judging 4o-Judge accuracy and GPT-4o-mini as the shared backbone for method comparisons.LoCoMo covers Single-hop, Multi-hop, Open-domain, and Temporal questions; additional metrics include Token F1, BLEU-4, ROUGE-1/2/L, and BertScore.Source: §4.1 · p. 5

4.2 Main Results

• CreaMem achieves the best performance on LoCoMo and LongMemEval-S across 4o-Judge accuracy, F1, BLEU-4, ROUGE-1/2/L, and BertScore.The strongest baselines are HippoRAG 2 on LoCoMo and MemGAS on LongMemEval-S.Source: Table 1 · p. 6, §4.2 · p. 7

4.4 Hyperparameter Analysis

• At k=30, 4o-Judge accuracy peaks and declines by k=40 as token cost grows nearly linearly; the same non-monotone pattern occurs for F1, ROUGE-1/2/L, and BertScore.The study adopts k=20 to match baseline token budgets, despite marginally higher LoCoMo accuracy at k=30.Source: Figure 2 · p. 6, §4.1 · p. 6, §4.4 · p. 8

4.2 Main Results

• Structured baselines retain a shared ceiling because unrelated scenes compete in one pool; CreaMem matches strong-baseline token use on LoCoMo and uses an order of magnitude fewer tokens than Full History on LongMemEval-S.The reported gains are attributed to memory organization and retrieval rather than a larger context budget.Source: §4.2 · p. 7

4.3 Ablation Study

• 2.3 points separate 3-scene from 1-scene under identical content and retrieval budget, while 3-scene provides the best balance among tested partition counts.Coarser partitions make unrelated scenes compete for retrieval slots, whereas finer partitions over-fragment memory and dilute per-memory information density.Source: §4.3 · p. 8
• 75.1% overall, 9.2 points above Episodic only, is achieved by Full System; its advantage exceeds 11 points on Temporal questions despite Episodic only using 42% more tokens.The comparison holds retrieval budget constant; a multi-hop case shows that removing either episodic or scene-level evidence can yield incomplete or temporally ungrounded answers.Source: Figure 3 · p. 7, Figure 4 · p. 7, §4.3 · p. 8

4.5 Architectural Validation

• A sample of 1,000 LoCoMo dialogue turns shows consistently high Life/Work/Interest coverage and strong agreement between GPT-4o and three human raters.The reported coverage gaps arise from turns difficult to assign to a scene, such as greetings or small talk.Source: §4.5 · p. 8, Table 2 · p. 8

4.6 Failure Analysis

• CreaMem’s remaining failures involve incomplete retrieval of linked entries and failure to combine retrieved evidence, motivating dialogue-turn links, segment fallback, and explicit evidence combination.In the examples, SeCom answers correctly when it retains a contiguous dialogue segment, while MemoryOS answers correctly when CreaMem fails to connect retrieved entries.Source: §4.6 · pp. 8–9

• CreaMem achieves consistent accuracy and multi-hop reasoning gains on LoCoMo and LongMemEval-S at comparable token cost.The conclusion attributes these gains to memory organization rather than supplying more context.Source: §5 · p. 9
• CreaMem partitions memory by users’ life scenes alongside Episodic and Core Memory, then combines dual-coded experience views through per-memory balanced sampling.The architecture is motivated by autobiographical memory organization and dual coding in cognitive psychology.Source: §5 · p. 9
• The conclusion suggests that organizing memory, rather than supplying more context, is the key lever for coherent long-term agents.Source: §5 · p. 9

Limitations

• The fixed Life/Work/Interest taxonomy is coarse and may not transfer uniformly across users, cultures, or specialized domains.Future work will investigate configurable or automatically induced scene partitions that adapt to individual users while retaining interpretable routing.Source: Limitations · p. 9
• End-to-end latency averages 2.9 seconds per query for CreaMem, compared with 1.98 seconds for A-Mem.A sufficiency check can trigger a second retrieval round, improving recall at the cost of additional latency and tokens; proposed mitigations include query-adaptive pruning, asynchronous retrieval, and learned stopping criteria.Source: Limitations · p. 9
• Evaluation remains challenging because write-time costs vary across methods, retrieval metrics assume direct alignment with reference evidence, and extraction depends on prompt and model choice.The passages identify batching granularity, module prompts, and stored-entry abstraction level as sources of write-time differences; standardized accounting, source-aligned retrieval evaluation, and method-agnostic extraction protocols remain future directions.Source: Limitations · p. 9

Category-level and open-weight evaluation

• CreaMem leads Overall, Multi-hop, and Temporal under matched GPT-4o-mini/GPT-4o, with 2.83- and 19.93-point strongest-baseline margins on Multi-hop and Temporal, and gains 5.97 points overall with Qwen3.6-35B-A3B.With Qwen3.6-35B-A3B, CreaMem also remains strongest on Multi-hop, Open-domain, and Temporal; SeCom leads Single-hop and MemoryOS leads Open-domain under the matched GPT-4o protocol, so dominance is not uniform.Source: A Supplementary Analyses · pp. 11–12

Routing, extraction, and taxonomy reliability

• 67.94% exact-match routing across 998 jointly judged LoCoMo turns accompanies 94.79% partial match and 74.53/73.65 Micro/Macro-F1.At 40% synthetic misrouting, QA falls 1.23 points, from 54.61% to 53.38%. In a 500-memory audit, Gemini 3.1 Pro/GPT-4o find 90.40/93.60% of entries retain a supported central fact, while 9.60/6.40% contain an unsupported central claim. Among 107 vertical-domain LongMemEval-S turns, exact routing is 68.22% and at-least-one-scene coverage is 99.03%; coverage is broad, though routing and extraction errors remain.Source: A Supplementary Analyses · p. 12

Allocation and architectural ablations

• 53.24% Full CreaMem exceeds 44.03% Episodic-only and 40.32–41.43% scene-only variants, while dynamic allocation improves only 0.26 points over balanced sampling.Dynamic allocation's 95% CI is [−0.78, 1.30] and adds 0.69 LLM calls per question; RRF changes overall accuracy by +0.06 and −0.91 points across two backbones, with both confidence intervals spanning zero. The analysis attributes the main gain to scene-aware dual encoding rather than allocation or fusion alone.Source: A Supplementary Analyses · p. 12

Representative failure cases

• Table 5 distinguishes failures to retrieve linked or implicit evidence from failures to combine evidence already present in context.These representative cases separate retrieval failures from evidence-combination failures.Source: A Supplementary Analyses · p. 12

QA and judge prompts

• For completeness, the QA and judge prompts reproduce those of MemGAS (Xu et al., 2026), with the judge accepting equivalent or complete answers and rejecting answers containing only a subset of required information.The listed key prompts are abridged for readability, and additional prompts are documented in the released code.Source: B Prompt Templates · pp. 12, 14, Figure 9 · p. 14, Figure 10 · p. 14

Retrieval planner

• The retrieval planner extracts 3–6 keywords and selects memory components based on the question, falling back to all four components and the question itself on parse failure or empty output.It routes event or timestamp questions to episodic memory and trait, habit, or preference questions to relevant Life Scene Memories; it can include Core Memory when a user profile is needed.Source: B Prompt Templates · p. 12

Scene Memory Extractor

• Life, Work, and Interest Memories share an extractor template whose scene definition varies, producing separate (content, importance_score) trait entries with importance_score ∈[0, 1].The scene definitions cover life matters such as health and family, work matters such as projects and career goals, and interests such as hobbies and travel.Source: B Prompt Templates · p. 14
• Extraction rules require distilled trait inferences rather than raw events, separate entries for distinct facts, verbatim preservation of concrete details, and exclusion of content outside the assigned scene.Entries are inserted by default and updated only when existing information has genuinely changed.Source: B Prompt Templates · p. 14