TraceMem: Narrative Memory for LLM Agents

Updated 4 July 2026

TraceMem is a framework that transforms fragmented dialogue snippets into coherent narrative memory schemata, ensuring long-term conversational coherence.
It employs a three-stage pipeline—short-term memory processing, synaptic consolidation, and systems consolidation—to structure and preserve multi-faceted user narratives.
TraceMem achieves state-of-the-art performance on the LoCoMo benchmark, demonstrating significant gains in multi-hop and temporal reasoning tasks.

TraceMem is a cognitively inspired framework for long-term conversational memory in LLM agents that weaves structured, narrative memory schemata from user conversational traces through a three-stage pipeline comprising Short-term Memory Processing, Synaptic Memory Consolidation, and Systems Memory Consolidation. It is designed for settings in which limited context windows and snippet-based memory systems fail to preserve the narrative coherence of prolonged dialogue histories, and it couples memory construction with an agentic search mechanism for downstream reasoning. Evaluation on the LoCoMo benchmark reports state-of-the-art performance, with especially strong gains on multi-hop and temporal reasoning, positioning TraceMem as a narrative-organization approach to conversational memory rather than a simple retrieval or context-extension technique (Shu et al., 10 Feb 2026).

1. Problem domain and conceptual basis

TraceMem addresses the long-term memory problem in dialogue systems: even large context windows do not guarantee effective use of long interaction histories, and raw context accumulation is vulnerable to phenomena such as “Lost-in-the-Middle.” The framework is motivated by the observation that many existing memory systems store dialogue as chunks, notes, or local summaries, so retrieval returns disjointed fragments rather than an integrated user-level account. TraceMem therefore treats the core failure mode as fragmentation: interactions are preserved as snippets, but the underlying narrative coherence of the dialogue stream is not captured (Shu et al., 10 Feb 2026).

The framework’s central construct is narrative coherence. In this setting, memory is not merely a bag of facts about a user, but a set of linked, temporally evolving threads that reflect hobbies, routines, family activities, health issues, and other recurrent developments across episodes. The architecture is explicitly cognitively inspired, drawing on synaptic consolidation, systems consolidation, episodic memory, semantic memory, and personal semantics. This suggests that TraceMem views conversational memory as an organizational problem as much as a storage problem: the challenge is not only to retain information, but to preserve the structure through which a user’s history becomes intelligible.

2. Short-term memory processing

The first stage, Short-term Memory Processing, performs deductive episodic segmentation and semantic extraction. Each utterance is classified as either TC (topic change) or TD (topic development) using surrounding context, and the decision is modeled as a context-dependent function $I_n = \mathcal{F}(D_n, \mathcal{D}_{pre}, D_{sub})$ . The resulting segmentation partitions the dialogue into contiguous topic-bounded episodes $E_1,\dots,E_k$ . In parallel, the framework extracts structured factual semantics for each utterance through

$s_n = \mathcal{T}(D_n \mid D_{n-1}, D_{n-2}, \text{Img}_n),$

where the current utterance is interpreted using recent textual context and associated image metadata when present. The paper describes both operations as XML-based structured prompting procedures rather than as trained segmentation or extraction models (Shu et al., 10 Feb 2026).

This stage produces two aligned outputs: a sequence of episodes $\mathcal{M}_{\text{epi}}$ and a sequence of semantic sets $\mathcal{M}_{\text{sem}}$ . The design is significant because TraceMem does not wait until retrieval time to decide what constitutes a meaningful memory unit. Topic boundaries are imposed early, and factual semantics are extracted before later summarization and clustering. A plausible implication is that later failures in long-horizon reasoning can be traced not only to retrieval quality but also to the quality of episode formation and semantic preservation at this first stage.

3. Synaptic and systems consolidation

The second stage, Synaptic Memory Consolidation, converts raw episodes into more durable user-centered traces. Each episode is summarized into an episodic memory $E_{s,i}$ , and then distilled jointly with the corresponding semantic set $S_i$ into a user-specific personal experience trace: $ET_i = \Phi(E_{s,i}, S_i; \mathcal{R}),$ where $\mathcal{R}$ is a rule-based filter that retains only the user’s own personal experiences and contextual details. At the stage level, synaptic consolidation maps episodes and semantic memories into episodic summaries and a set $U$ of distilled personal experience traces. This is the bridge between raw event memory and long-term autobiographical structure (Shu et al., 10 Feb 2026).

The third stage, Systems Memory Consolidation, organizes those traces into higher-order narrative structure through two-stage hierarchical clustering. The trace embeddings are represented as

$E_1,\dots,E_k$ 0

The implementation uses PCA for dimensionality reduction, UMAP for manifold learning, HDBSCAN for density-based clustering, and KNN to reassign low-density noise points. Topic-level clustering uses UMAP n_neighbors = 10 and HDBSCAN min_cluster_size = 5; thread-level clustering uses n_neighbors = 2 and min_cluster_size = 2. Text embeddings are generated with text-embedding-3-small and stored in ChromaDB (Shu et al., 10 Feb 2026).

The outcome is a set of user memory cards encoding narrative memory schemata. Each card has a three-level structure: a top-level theme title, middle-level topical sections, and bottom-level thread entries. Each thread entry contains a descriptive title, a coherent summary, and a unique identifier. Example card themes include “Melanie’s Journey of Self-Care, Family Activities, and Community Involvement” and “Melanie’s Artistic Journey: Painting, Pottery, and Personal Growth.” The representational move is consequential: TraceMem does not stop at episodic summaries, but further reorganizes them into coherent, time-evolving narrative threads under unifying themes.

4. Memory utilization and agentic search

At inference time, TraceMem uses a three-step Agentic Search process rather than one-shot similarity retrieval. First, the incoming query retrieves the top- $E_1,\dots,E_k$ 1 episodic memories from the global vector database, with $E_1,\dots,E_k$ 2. Second, the system selects relevant memory cards using user cues or topic cues in the query. Third, it inspects the selected cards—especially topic titles, thread titles, thread summaries, and thread identifiers—to pinpoint relevant narrative threads and fetch the corresponding thread content directly from the vector store (Shu et al., 10 Feb 2026).

This yields a composite retrieval context: discrete episodic memories provide local grounding, while narrative threads provide user-level continuity. The paper characterizes this fusion as a “multi-dimensional contextual foundation” for reasoning and response generation. TraceMem therefore uses two complementary evidence forms at answer time: event-level memory and narrative-level schema. This suggests that the framework treats long-term conversational competence as requiring both source episodes and an organized model of the user’s ongoing story.

5. Empirical performance on LoCoMo

TraceMem is evaluated on LoCoMo, which contains 10 long dialogues, about 600 turns per dialogue on average, 1,540 annotated questions, and four reasoning categories: Single-hop, Multi-hop, Temporal, and Open-domain. The main experiments use GPT-4o-mini and GPT-4.1-mini as backbones, with accuracy judged by GPT-4o-mini using the Nemori prompt template. Against FullText, NaiveRAG, A-Mem, LightMem, and Nemori, TraceMem attains the highest overall scores on both backbones (Shu et al., 10 Feb 2026).

Method	Overall (GPT-4o-mini)	Overall (GPT-4.1-mini)
FullText	0.7305	0.8740
NaiveRAG	0.4623	0.5188
A-Mem	0.4130	0.5955
LightMem	0.7007	0.7669
Nemori	0.7403	0.8052
TraceMem	0.9019	0.9227

The category-level results are particularly notable in the reasoning regimes that most depend on structured long-term memory. With GPT-4o-mini, TraceMem reaches 0.9310 on SingleHop, 0.9220 on MultiHop, 0.8660 on Temporal, and 0.7083 on OpenDomain. With GPT-4.1-mini, it reaches 0.9512, 0.8936, 0.9097, and 0.8021 respectively. The contrast with the strongest baseline is especially sharp on multi-hop and temporal questions, which the paper interprets as evidence that coherent narrative structure helps with integrating scattered evidence and preserving chronology (Shu et al., 10 Feb 2026).

Ablation results reinforce that interpretation. “Episodic Memory Only” achieves an overall score of 0.8429; adding semantic representations raises this to 0.8701 with 20 semantic items and 0.8792 with 40; removing agentic search yields 0.8844; the complete TraceMem system reaches 0.9019. The gains are strongest on MultiHop and Temporal categories. Memory construction statistics also show substantial structural transformation. Under GPT-4o-mini, the system produced 925 episodes, 1379 experiences, 475 threads, with Discard: 235.5 and Discard Rate: 25.46%. Under GPT-4.1-mini, the counts were 1212, 2104, 736, 160, and 13.20%. The paper defines

$E_1,\dots,E_k$ 3

and interprets the GPT-4.1-mini regime as more “conservative.”

6. Interpretation, limitations, and neighboring research

TraceMem’s open discussion identifies several issues that remain unresolved within the current framework: memory rewriting upon retrieval, regulated forgetting, selective retrieval to avoid interference, and personality-dependent memory prioritization. The paper also leaves several engineering details underspecified in the main text, including the exact card-update procedure over time, the full prompt contents, and the latency or token-cost implications of repeated consolidation and agentic retrieval. Evaluation is confined to LoCoMo, so generalization to other long-term settings remains an open question (Shu et al., 10 Feb 2026).

Within the broader memory-systems literature, TraceMem occupies a specific niche. It is distinct from the diagnostic MemTrace framework that turns memory pipelines into execution graphs for error tracing and attribution in LLM memory systems (Deng et al., 27 May 2026). It also differs from the benchmark MemTrace that evaluates long-term memory at the level of a “knowledge point” under controlled variations of age, question type, and evidence condition (Long et al., 15 Jun 2026). In another domain, AndroTMem studies long-horizon Android GUI agents and proposes Anchored State Memory, where memory units are sparse, causally linked intermediate-state anchors rather than narrative schemata (Shi et al., 19 Mar 2026). Selective Memory Retention for Long-Horizon LLM Agents isolates bounded retention policies over episodic trajectory summaries and shows that selective retention matters most under noisy-write conditions (Reddy, 28 Jun 2026). MEMTRACK shifts evaluation toward multi-platform enterprise workflows with asynchronous traces across Slack, Linear, Git, and file systems, emphasizing acquisition, selection, and conflict resolution (Deshpande et al., 1 Oct 2025). MemMark addresses provenance rather than retrieval quality by watermarking latent memory-write decisions so attribution can survive snapshot-only settings (Zhang et al., 24 May 2026). This suggests that TraceMem is best understood as a narrative-organization layer within a wider research landscape that also includes attribution, benchmarking, anchored interaction memory, and retention control.