Source-linked AI summary
SimpleMem: Efficient Lifelong Memory for LLM Agents
Jiaqi Liu, Yaofeng Su, Peng Xia, Siwei Han, Zeyu Zheng, Cihang Xie, Mingyu Ding, Huaxiu Yao
TL;DR
LLM agents need efficient memory for long-term interaction, while existing systems retain redundant histories or incur high costs filtering them. SimpleMem applies semantic structured compression, online synthesis, and intent-aware retrieval planning to organize and retrieve compact memories. It consistently outperforms existing memory systems, including a 26.4% F1 improvement on LoCoMo and up to 30-fold lower inference-time token consumption.
Problem
Existing memory systems either preserve complete interaction histories containing substantial redundancy or require costly filtering, limiting efficient long-term memory use.
Method
SimpleMem uses semantic structured compression, online semantic synthesis, and intent-aware retrieval planning to compress, organize, and adaptively retrieve agent memories.
Results
26.4% F1 improvement over Mem0 on LoCoMo and roughly 30× lower token consumption than full-context models are reported, while SimpleMem consistently outperforms existing memory systems across LoCoMo and LongMemEval-S.
Takeaways & Limitations
SimpleMem demonstrates a strong balance between retrieval accuracy and inference efficiency across benchmark datasets and model scales.
Abstract
from arXiv · showhide
To support long-term interaction in complex environments, LLM agents require memory systems that manage historical experiences. Existing approaches either retain full interaction histories via passive context extension, leading to substantial redundancy, or rely on iterative reasoning to filter noise, incurring high token costs. To address this challenge, we introduce SimpleMem, an efficient memory framework based on semantic lossless compression. We propose a three-stage pipeline designed to maximize information density and token utilization: (1) Semantic Structured Compression, which distills unstructured interactions into compact, multi-view indexed memory units; (2) Online Semantic Synthesis, an intra-session process that instantly integrates related context into unified abstract representations to eliminate redundancy; and (3) Intent-Aware Retrieval Planning, which infers search intent to dynamically determine retrieval scope and construct precise context efficiently. Experiments on benchmark datasets show that our method consistently outperforms baseline approaches in accuracy, retrieval efficiency, and inference cost, achieving an average F1 improvement of 26.4% in LoCoMo while reducing inference-time token consumption by up to 30-fold, demonstrating a superior balance between performance and efficiency. Code is available at https://github.com/aiming-lab/SimpleMem.
1. Introduction
LLM agents need memory systems for reliable long-term interaction, but existing approaches either preserve redundant histories or use costly filtering. SimpleMem addresses this with structured semantic compression and reports higher accuracy with much lower token use.
- Long-term, multi-turn interactions require memory systems that efficiently manage and use historical experience.
- Full-context extension accumulates repetitive, low-entropy dialogue, reducing memory information density and harming retrieval and downstream reasoning.
- SimpleMem uses structured semantic compression to improve information efficiency under fixed context and token budgets.
- SimpleMem combines semantic compression, online synthesis, and intent-aware retrieval planning to organize and retrieve memory efficiently.
2. The SimpleMem Architecture
SimpleMem transforms dialogue into compact, structured memories, consolidates related observations during writing, and plans retrieval according to inferred intent. Its architecture combines semantic, lexical, and symbolic access to construct compact contexts under token constraints.
- Semantic Structured Compression: Sliding windows provide fixed-length processing units, while an implicit semantic gate preserves high-entropy content and discards low-density windows without explicit threshold tuning.
- Semantic Structured Compression: A unified transformation jointly extracts facts, resolves coreference, anchors time, and produces self-contained memory units from each dialogue window.
- Semantic Structured Compression: Semantic Structured Compression filters low-utility dialogue and converts informative windows into compact, context-independent memory units.
- Structured Multi-View Indexing: Each memory unit receives complementary semantic, lexical, and symbolic representations, supporting conceptual similarity, exact keyword matching, and metadata filtering.
- Online Semantic Synthesis: Online Semantic Synthesis merges related session observations into unified, high-density entries before database commitment, reducing fragmentation and redundancy.
- Intent-Aware Retrieval Planning: Intent-Aware Retrieval Planning infers latent search intent to adapt retrieval scope and depth, then queries all index views in parallel and deduplicates results.
- Intent-Aware Retrieval Planning: The resulting context preserves relevance signals from multiple views while scaling retrieval volume with inferred query complexity.
3. Experiments
SimpleMem is evaluated on long-term conversational benchmarks across model scales, where it consistently improves accuracy while reducing retrieval and inference costs. Component analyses attribute these gains to semantic compression, online synthesis, and intent-aware retrieval planning.
- Overall performance: SimpleMem consistently outperforms existing memory systems across LoCoMo and LongMemEval-S model scales.It leads the reported comparisons on both benchmarks, including high-capability and smaller models.
- LoCoMo results: 43.24 Average F1 on GPT-4.1-mini LoCoMo exceeds Mem0 (34.20) and the full-context baseline (18.70).SimpleMem also reaches 58.62 F1 in Temporal Reasoning versus 48.91 for Mem0.
- LongMemEval-S results: 76.87% average accuracy on LongMemEval-S with gpt-4.1-mini surpasses LightMem (68.67%), Mem0 (59.81%), and full-context (39.57%).In Multi-Session evaluation, SimpleMem achieves 60.92% accuracy versus 47.37% for LightMem and 30.08% for full-context.
- Efficiency analysis: 530-580 tokens per query reduce inference usage by roughly 30× versus full-context approaches and 40-50% versus Mem0 and A-Mem.On GPT-4.1-mini, SimpleMem uses 531 tokens, compared with 643 for ReadAgent, while achieving higher accuracy.
- Smaller models: 33.45 Average F1 on Qwen3-8b surpasses Mem0 (25.80) and LightMem (22.23), while SimpleMem also improves results on 3B and 1.5B models.Qwen2.5-3b reaches 17.98 F1 with SimpleMem versus 13.03 with Mem0.
- Component analysis: Ablations show that removing semantic compression reduces Temporal F1 by 56.7%, disabling online synthesis lowers multi-hop performance by 31.3%, and fixed-depth retrieval causes 26.6% and 19.4% drops on open-domain and single-hop tasks.The reported degradations are associated with unresolved temporal context, fragmented evidence, and poorly matched retrieval scope.
4. Related Work
Prior memory systems extend context or impose structure, but often retain redundant logs and incur retrieval or processing costs. SimpleMem addresses these limitations with more efficient retrieval for dynamic episodic memory.
- Virtual-context systems extend interaction length through paging or stream controllers but typically store raw logs, increasing redundancy and processing costs.
- Structured and graph-based memory systems improve coherence through structural priors but may still rely on raw or minimally processed content.
- Long-context approaches suffer from degraded reasoning as context length increases and impose substantial computational overhead for lifelong agents.
- RAG methods decouple memory from inference but are largely optimized for static knowledge bases, limiting effectiveness for dynamic, time-sensitive episodic memory.
- SimpleMem improves retrieval efficiency through intent-aware planning for dynamic episodic memory.
5. Conclusion
SimpleMem is an agent memory architecture based on semantic lossless compression and evaluated on LoCoMo and LongMemEval-S. Its design treats memory as an active process involving compression, synthesis, and adaptive retrieval.
- SimpleMem treats memory as an active process rather than passive storage.
- The architecture combines Semantic Structured Compression, Online Semantic Synthesis, and Intent-Aware Retrieval Planning.
- SimpleMem was empirically evaluated on the LoCoMo and LongMemEval-S benchmarks.
A. Detailed System Prompts
The appendix specifies prompts for compressing dialogue, planning retrieval, synthesizing answers, and evaluating LongMemEval responses. These prompts emphasize compact memories, adaptive search, temporal consistency, and semantic correctness.
- A. Detailed System Prompts: The reproducibility appendix provides model-agnostic prompts optimized for GPT-4o-mini experiments.
- A.1. Stage 1: Semantic Structured Compression Prompt: Semantic Structured Compression transforms raw dialogue windows into compact, context-independent memory units while excluding low-information content.
- A.1. Stage 1: Semantic Structured Compression Prompt: Stage 1 discards social filler and redundant confirmations, resolves references, normalizes time, and extracts minimal factual statements.
- A.1. Stage 1: Semantic Structured Compression Prompt: The compression prompt receives a dialogue window, window start time, and participant list, then outputs structured memory content with entities, topics, timestamps, and salience.
- A.2. Stage 2: Adaptive Retrieval Planning Prompt: Adaptive Retrieval Planning estimates query complexity and generates a structured plan that minimizes unnecessary context usage.
- A.2. Stage 2: Adaptive Retrieval Planning Prompt: The retrieval plan distinguishes direct lookups from multi-event synthesis and combines lexical, temporal, and semantic signals.
- A.3. Stage 3: Reconstructive Synthesis Prompt: Reconstructive Synthesis combines abstract representations with detailed memory units, preserving temporal consistency and grounding answers in retrieved memory.
- A.4. LongMemEval Evaluation Prompt: The LongMemEval judge labels answers CORRECT or WRONG using semantic and temporal alignment, accepting equivalent dates despite formatting or relative-time differences.
B.1. Dataset Description
LoCoMo evaluates long-term conversational reasoning across lengthy, complex dialogues, while LongMemEval-S tests precise answer localization in exceptionally long histories. Together, they target temporal, multi-hop, open-domain, and direct retrieval capabilities.
- LoCoMo contains conversations of 200–400 turns with complex temporal shifts and interleaved topics.
- LoCoMo tests multi-hop reasoning, temporal reasoning, open-domain questions, and single-hop fact retrieval.
- LongMemEval-S requires precise localization of answers across temporal events, user preferences, and other categories within exceptionally long histories.
B.2. Hyperparameter Configuration
The reported hyperparameters were selected to balance memory compactness with retrieval recall, focusing on thresholds for semantic structured compression and recursive consolidation.
- Hyperparameters balance memory compactness and retrieval recall.These values were used for the results reported in Section 3.
- Thresholds govern semantic structured compression and recursive consolidation.
- Table 7 summarizes the configuration used to obtain Section 3 results.
B.3. Hyperparameter Sensitivity Analysis
SimpleMem reaches near-peak performance with shallow retrieval and remains stable as retrieval depth increases. The sensitivity analysis supports adaptive retrieval while emphasizing the information density of its memory units.
- 35.20 F1 is achieved with a single retrieved entry, and approximately 99% of peak performance is reached at k = 3.The analysis uses average F1 on LoCoMo with the GPT-4.1-mini backend.
- At k = 3, SimpleMem scores 42.85 compared with its peak of 43.45 at k = 10.The table labels this pattern rapid saturation and associates it with Atomic Entries.
- SimpleMem maintains stable accuracy when retrieving up to 20 entries, unlike MemGPT, which degrades at larger k.
- Adaptive retrieval can expand context for complex reasoning tasks without introducing excessive irrelevant information.This follows from SimpleMem's robustness to increased retrieval depth.
- The configuration uses adaptive thresholds to balance memory compactness and retrieval effectiveness.