Research map / Multi-agent systems and games / LLM agents
A-Mem: Agentic Memory for LLM Agents
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 A-Mem, a Zettelkasten-inspired agentic memory system in which LLM-generated notes are dynamically linked and evolved, evaluated on LoCoMo and DialSim across several LLM backbones.
Problem
LLM agents need long-term memory, but existing memory systems rely on predefined storage structures, fixed workflows, and rigid schemas (e.g., graph databases), limiting adaptability across tasks.
Approach
Each memory is stored as an LLM-generated structured note (keywords, tags, contextual description, embedding); new memories trigger LLM-driven link generation to top-k neighbors and memory evolution that rewrites the attributes of existing notes, following Zettelkasten principles.
Key results
- On LoCoMo with GPT-4o-mini, A-Mem reaches 27.02 F1 on Multi Hop vs 25.02 for LoCoMo and 26.65 for MemGPT, and uses about 1,200-2,500 tokens versus ~16,900 for LoCoMo/MemGPT.
- On DialSim, A-Mem achieves F1 3.45 versus 2.55 for LoCoMo and 1.18 for MemGPT.
- Ablation on GPT-4o-mini shows removing link generation and evolution drops Multi Hop F1 from 27.02 to 9.65.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use A-Mem instead of LoCoMo | GPT-4o-mini · DialSim | F1: 2.55 -> 3.45 | Helped |
| Use A-Mem instead of LoCoMo | GPT-4o-mini · DialSim | BLEU-1: 3.13 -> 3.37 | Helped |
| Use A-Mem instead of LoCoMo | GPT-4o-mini · LoCoMo Temporal | F1: 18.41 -> 45.85 | Helped |
| Use A-Mem instead of LoCoMo | GPT-4o-mini · LoCoMo Adversarial | F1: 69.23 -> 50.03 | Hurt |
| Add link generation and memory evolution (ablation) | A-Mem · LoCoMo Multi Hop | F1: 9.65 -> 27.02 | Helped |
| Add link generation and memory evolution (ablation) | A-Mem · LoCoMo Temporal | F1: 24.55 -> 45.85 | Helped |
| Add link generation and memory evolution (ablation) | A-Mem · LoCoMo Open Domain | F1: 7.77 -> 12.14 | Helped |
| Add link generation and memory evolution (ablation) | A-Mem · LoCoMo Single Hop | F1: 13.28 -> 44.65 | Helped |
| Add link generation and memory evolution (ablation) | A-Mem · LoCoMo Adversarial | F1: 15.32 -> 50.03 | Helped |
| Add memory evolution (ablation) | A-Mem with link generation · LoCoMo Multi Hop | F1: 21.35 -> 27.02 | Helped |
| Add memory evolution (ablation) | A-Mem with link generation · LoCoMo Temporal | F1: 31.24 -> 45.85 | Helped |
| Increase memory retrieval parameter k | A-Mem with GPT-4o-mini · LoCoMo task categories | performance | Mixed |
Surprising findings
- On GPT-4o, LoCoMo and MemGPT outperform A-Mem on Open Domain and Adversarial categories, and LoCoMo beats A-Mem on Single Hop F1 (61.56 vs 48.43).
- Retrieval time for A-Mem stays sub-microsecond to a few microseconds at 1M memories, but MemoryBank is faster, and ReadAgent is much slower.
Limitations
- Memory organization quality depends on the underlying LLM, which may produce inconsistent descriptions or links.
- Only text-based interactions are handled; multimodal memory is left to future work.
- Evaluation is limited to QA on LoCoMo and DialSim; results on GPT-4o are mixed relative to baselines.
Experimental setup
Datasets: LoCoMo DialSim
Benchmarks: LoCoMo DialSim
Metrics: F1 BLEU-1 ROUGE-L ROUGE-2 METEOR SBERT Similarity token length retrieval time memory usage
Baselines: LoCoMo ReadAgent MemoryBank MemGPT
Builds on: Zettelkasten method MemGPT MemoryBank ReadAgent LoCoMo
Closest papers by meaning
- Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.95)
- AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications (ICML 2026 · similarity 0.93)
- EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective (NeurIPS 2026 · similarity 0.93)
- PlugMem: A Task-Agnostic Plugin Memory Module for LLM Agents (ICML 2026 · similarity 0.92)
- G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems (NeurIPS 2025 · similarity 0.91)
- UMEM: Unified Memory Extraction and Management Framework for Generalizable Memory (ICML 2026 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.