Research map / Large language models / AI and society
MemoryBank: Enhancing Large Language Models with Long-Term Memory
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 MemoryBank, a long-term memory mechanism for LLMs with Ebbinghaus-inspired memory updating, instantiated as the SiliconFriend AI companion built on ChatGPT, ChatGLM and BELLE.
Problem
LLMs lack long-term memory, which hinders sustained interaction scenarios such as personal companionship, psychological counseling, and secretarial assistance.
Approach
A unified memory storage (conversations, hierarchical event summaries, user portraits) with dense retrieval over FAISS and a memory-strength update rule modeled on the Ebbinghaus forgetting curve, so memories decay with time and are reinforced when recalled; the memory is injected into the prompt of any LLM backbone.
Key results
- SiliconFriendChatGPT achieved the best English results with correctness 0.716, coherence 0.912 and ranking score 0.818.
- SiliconFriendBELLE reached the highest Chinese retrieval accuracy (0.856) and correctness 0.603 among Chinese variants.
- ChatGLM and BELLE variants showed high retrieval accuracy (0.809 to 0.856), supporting the claim that MemoryBank generalizes across open and closed LLMs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ChatGPT rather than the other SiliconFriend variants | SiliconFriend · English simulated dialogs | response correctness: 0.479, 0.438 -> 0.716 | Helped |
| Use ChatGPT rather than the other SiliconFriend variants | SiliconFriend · Chinese simulated dialogs | response correctness: 0.603 -> 0.655 | Helped |
| Use ChatGPT rather than the other SiliconFriend variants | SiliconFriend · Chinese simulated dialogs | memory retrieval accuracy: 0.856 -> 0.711 | Hurt |
| Use BELLE rather than ChatGLM | SiliconFriend · Chinese simulated dialogs | response correctness: 0.418 -> 0.603 | Helped |
| Use ChatGLM rather than ChatGPT | SiliconFriend · Chinese simulated dialogs | memory retrieval accuracy: 0.711 -> 0.84 | Helped |
Surprising findings
- Open-source SiliconFriend variants retrieve memories well but lag ChatGPT on response correctness and coherence, attributed to weaker base models.
Limitations
- The memory updating model is explicitly described as exploratory and highly simplified (discrete strength, ignores overlearning and meaningful material effects).
- Quantitative evaluation uses a small simulated set (15 users, 194 probing questions) scored by human annotators, with no public benchmark release.
- Open-source variants underperform ChatGPT on correctness and coherence.
Experimental setup
Datasets: 38k psychological dialogues from online psychological QA websites simulated 10-day multi-user dialogue memory storage generated by ChatGPT (15 virtual users, 450 topics)
Benchmarks: 194 manually written memory probing questions (97 English, 97 Chinese) over simulated dialogs
Metrics: memory retrieval accuracy response correctness contextual coherence model ranking score
Baselines: ChatGLM ChatGPT BELLE
Builds on: Dense Passage Retrieval LoRA Ebbinghaus forgetting curve
Closest papers by meaning
- LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory (ICLR 2025 · similarity 0.91)
- PersonaVLM: Long-Term Personalized Multimodal LLMs (CVPR 2026 · similarity 0.88)
- AMemGym: Interactive Memory Benchmarking for Assistants in Long-Horizon Conversations (ICLR 2026 · similarity 0.88)
- Mem-PAL: Towards Memory-based Personalized Dialogue Assistants for Long-term User-Agent Interaction (AAAI 2026 · similarity 0.87)
- Personalize-then-Store: Benchmarking and Learning Personalized Memory for Long-Horizon Agents (NeurIPS 2026 · similarity 0.87)
- SeCom: On Memory Construction and Retrieval for Personalized Conversational Agents (ICLR 2025 · similarity 0.87)
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.