Technical Summary: DimMem – Dimensional Structuring for Efficient Long-Term Agent Memory

1. Problem Statement

Large Language Model (LLM) agents require long-term memory to function as persistent assistants capable of recalling user preferences, past decisions, and evolving facts across multiple sessions. Existing memory systems face a fundamental fidelity–efficiency trade-off:

• Raw Dialogue Histories: Preserving full context ensures fidelity but is computationally expensive to store, retrieve, update, and inject into the model context.
• Flat Facts/Summaries: These are efficient but often discard the temporal, causal, intentional, and contextual structures necessary for precise recall (e.g., when an event happened, why a plan changed, or what the assistant previously recommended).
• Complex Structures: Graph-based or hierarchical systems recover relational structure but introduce significant construction and maintenance overhead.
• Learned Policies: Approaches that learn memory operations (CRUD) via reinforcement learning or fine-tuning often struggle because the underlying memory representation lacks explicit structural dimensions (time, reason, purpose), forcing the policy to infer missing information from impoverished records.

The core challenge is that queries often depend on specific dimensions (temporal, causal, intentional) rather than just semantic similarity. Current systems fail to expose these dimensions explicitly at the storage level, leading to retrieval errors and inefficient context usage.

2. Methodology: DimMem Framework

DimMem proposes a lightweight dimensional memory framework where every memory is represented as an atomic, typed, and self-contained unit with explicit dimensional fields.

2.1 Dimensional Memory Representation

Instead of storing raw text or flat summaries, DimMem converts dialogue streams into structured records:
$$m_i = \langle id_i, C_i, T_i, D_i, \rho_i \rangle$$

• $C_i$ (Content): A self-contained textual memory where pronouns are resolved and relative times are normalized.
• $T_i$ (Type): Explicit classification into fact (stable knowledge), episodic (specific events), or profile (long-term personal semantics).
• $D_i$ (Dimensions): A decoupled set of explicit fields:
  • Time: Validity or occurrence time.
  • Location: Physical or contextual location.
  • Reason: Causal background.
  • Purpose: Communicative or task intent.
  • Keywords: Entity and topic anchors.
• $\rho_i$ (Pointer): A reference to the source window/turn, allowing retrieval of original assistant replies on demand without storing them in every record.

2.2 Overlap-Aware Extraction

To handle long dialogues, the system segments the conversation into overlapping windows. Memories are extracted such that:

• Overlapping tokens are used to resolve coreference and normalize time.
• New memories are grounded strictly in the non-overlapping suffix of the window to prevent duplication.
• A compact extractor (fine-tuned via SFT) maps these windows to the dimensional schema.

2.3 Dimension-Aware Operations

The framework unifies retrieval, update, and assistant recall through the dimensional schema:

• Tri-Path Retrieval: Queries are parsed into the same schema. Retrieval combines:
  1. Lexical Matching ($R_{bm25}$): For exact entities.
  2. Dense Retrieval ($R_{dense}$): For semantic paraphrases.
  3. Dimension Matching ($R_{dim}$): Explicit scoring based on agreement between query constraints (time, type, location, reason) and memory dimensions. This prevents semantically similar but temporally or causally mismatched memories from dominating.
• Dynamic Assistant Recall: The system predicts if a query requires previous assistant context ($I_{ast}$). If so, it retrieves the specific source reply via $\rho_i$; otherwise, it relies solely on structured user memories. This avoids inflating the context for queries that do not need it.
• Dimension-Aware Update: New records are filtered by memory type and keyword overlap before expensive embedding comparison or LLM consolidation. This reduces the candidate set for merging, superseding, or keeping both records, ensuring that only structurally plausible conflicts trigger LLM decisions.

2.4 Learnable Structured Extraction

The paper demonstrates that high-quality dimensional extraction is a learnable capability. By fine-tuning a compact model (Qwen3-4B) on synthetic data generated by a teacher model (GPT-5.4) using the DimMem schema, the system achieves performance comparable to much larger extractors. This shifts the burden from "prompting large models" to "training compact models on a structured schema."

3. Key Contributions

1. Dimensional Memory Representation: A shift from flat passages or isolated facts to atomic, typed records with explicit dimensions (time, location, reason, purpose, keywords) that serve as the operational interface for the memory system.
2. Dimension-Aware Operations: Mechanisms for retrieval, update, and assistant recall that align parsed query intents with explicit memory dimensions, improving temporal, causal, and multi-session reasoning while reducing costs.
3. Learnable Structured Extraction: Empirical evidence that compact models (e.g., 4B parameters) can learn to construct high-quality dimensional memories, challenging the notion that such structured extraction requires proprietary or massive-scale models.

4. Experimental Results

The framework was evaluated on LoCoMo-10 and LongMemEval-S benchmarks.

• Performance: DimMem achieved 81.43% overall accuracy on LoCoMo-10 and 78.20% on LongMemEval-S, outperforming existing lightweight systems (e.g., LightMem, SimpleMem, MemOS).
• Efficiency: It reduced the per-query token cost on LoCoMo by 24% compared to baselines.
• Compact Extractor: A fine-tuned Qwen3-4B extractor surpassed LightMem (using GPT-4.1-mini) on both benchmarks and reached performance comparable to a Qwen3-30B-A3B extractor.
• Ablation Studies:
  • Removing dimensional structure (Content-Only) caused the largest performance drops, particularly in temporal and multi-hop reasoning.
  • Removing the dimension-based retrieval route reduced precision, confirming that lexical and dense retrieval alone cannot enforce strict constraints.
  • Dynamic Assistant Recall: Triggered for only 12.2% of queries but improved assistant-dependent accuracy by 48.2% on specific sub-tasks, reducing total token consumption by 65.6%.
  • Update Efficiency: Dimensional filtering reduced embedding computations by 75.5% (LongMemEval) and 59.2% (LoCoMo) while identifying more valid merges than embedding-only baselines.

5. Significance and Claims

The paper posits that explicit dimensional structuring is an effective and efficient foundation for long-term memory in LLM agents. The significance lies in:

• Inductive Bias: Treating time, reason, purpose, and type as first-class citizens in the memory schema acts as a strong inductive bias, directly improving retrieval precision and reasoning capabilities.
• Decoupling Control from Representation: While prior work focused on learning when to update memory, DimMem focuses on how to construct the memory substrate. It argues that a high-quality, structured substrate is a prerequisite for any effective memory policy.
• Scalability: The ability to train compact models to perform structured extraction suggests that sophisticated memory management does not inherently require massive computational resources, making long-term agent memory more accessible and deployable.

The authors conclude that DimMem offers a middle path between the high cost of raw context and the low fidelity of flat summaries, demonstrating that structured, dimension-aware memory is both performant and learnable by smaller models.