Research map / Large language models / Reasoning and chain of thought

SimpleMem: Efficient Lifelong Memory for LLM Agents

ICML 2026 · Large language models › Reasoning and chain of thought · approach: Efficient LLM serving · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SimpleMem, a LLM-agent memory framework with LLM-based semantic compression, intra-session online synthesis, and intent-aware multi-view retrieval, reporting higher F1 on LoCoMo and LongMemEval-S at ~30x fewer tokens than full-context.

Problem

LLM agents need long-term memory, but full-history retention causes redundancy and token cost, while iterative reasoning-based filtering is expensive.

Approach

Treats memory writing as LLM-driven semantic lossless compression: windows are converted into self-contained, coreference- and time-normalized atomic facts, synthesized on the fly within a session, and retrieved via a planner that issues dense, BM25, and metadata queries in parallel and unions the results.

Key results

• On LoCoMo with GPT-4.1-mini, SimpleMem reaches 43.24 average F1 versus 34.20 for Mem0, using about 531 tokens per query versus about 16,910 for full context.
• On LongMemEval-S with GPT-4.1, SimpleMem reaches 83.97% average accuracy versus 58.51% for Mem0.
• On Qwen2.5-3B, SimpleMem achieves 17.98 average F1 versus 13.03 for Mem0 with the same backbone.

Surprising findings

• Removing semantic compression lowers temporal F1 from 58.62 to 25.40, a 56.7% drop, the largest single ablation effect.
• Qwen2.5-1.5B with SimpleMem (25.23 F1) outperforms Qwen3-1.7B with Mem0 (21.19 F1).

Limitations

• Relies on the LLM to perform compression and planning, so extraction quality depends on backbone and prompts; prompts were optimized for GPT-4o-mini.
• Ablation and efficiency results are reported with a single backend (GPT-4.1-mini) and the evaluation covers only two benchmarks.
• Reported LoCoMo numbers for some baselines and the abstract claims are not fully consistent with the tables (e.g., 26.4% F1 gain and 30x token reduction vary across settings).

Experimental setup

Datasets: LoCoMo LongMemEval-S

Benchmarks: LoCoMo LongMemEval-S

Metrics: F1 BLEU-1 accuracy token cost construction time retrieval time

Baselines: LoCoMo ReadAgent MemoryBank MemGPT A-Mem LightMem Mem0 Full-context

Builds on: MemGPT Mem0 A-Mem LightMem Complementary Learning Systems

Closest papers by meaning

• MemForest: Efficient Agent Memory Management via EventTree Partitioning and Progressive Merging (NeurIPS 2026 · similarity 0.91)
• A-Mem: Agentic Memory for LLM Agents (NeurIPS 2025 · similarity 0.91)
• Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.90)
• AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications (ICML 2026 · similarity 0.89)
• Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents (ICML 2026 · similarity 0.89)
• Beyond RAG for Agent Memory: Retrieval by Decoupling and Aggregation (NeurIPS 2026 · similarity 0.89)