AuthorsWanjun Zhong, Lianghong Guo, Qiqi Gao et al.
TL;DR
MemoryBank adds Ebbinghaus-style memory updating on top of long-term storage and retrieval, enabling SiliconFriend to reach 0.716 correctness and 0.912 coherence on English probing questions.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
MemoryBank targets the limitation that LLMs lack long-term memory, which is critical for personal companions, psychological counseling, and secretarial assistance.
Without persistent memory, systems like SiliconFriend cannot recall past conversations or understand user personality, degrading rapport, emotional support, and personalized task management over time.
HOW IT WORKS
MemoryBank’s core mechanism links Memory Storage, Memory Retrieval, and a Memory Updating Mechanism to store conversations, event summaries, and evolving user portraits.
You can think of MemoryBank like a human brain with a diary: detailed logs as notes, summaries as memories, and Ebbinghaus-style rehearsal deciding what to forget or reinforce.
This design lets MemoryBank keep rich user history and personality outside the context window, enabling SiliconFriend to recall and adapt in ways a plain prompt buffer cannot.
DIAGRAM
This diagram shows how MemoryBank retrieves relevant memories and user portraits for a new SiliconFriend query and injects them into the meta prompt.
DIAGRAM
This diagram shows how MemoryBank is evaluated using 10-day multi-user simulated dialogs and 194 probing questions scored by human annotators.
PROCESS
Memory Storage
MemoryBank logs daily multi-turn conversations with timestamps and builds hierarchical Event Summary and User Portrait layers for each user.
Memory Retrieval
MemoryBank encodes the current context and searches pre-encoded memories with a dual-tower dense retriever and FAISS to find relevant pieces.
Memory Updating Mechanism
MemoryBank updates memory strength S using the Ebbinghaus Forgetting Curve, increasing S and resetting time t when a memory is recalled.
SiliconFriend Integration
MemoryBank feeds relevant memory, global event summary, and global user portrait into SiliconFriend’s meta prompt to generate personalized responses.
KEY CONTRIBUTIONS
MemoryBank long-term memory mechanism
MemoryBank introduces Memory Storage, Memory Retrieval, and a Memory Updating Mechanism so LLMs can store, recall, and update long-term memories and user portraits.
SiliconFriend AI companion
MemoryBank powers SiliconFriend, an AI companion tuned with 38k psychological conversations, enabling empathetic responses and personality-aware suggestions in English and Chinese.
Generalizable across LLMs and languages
MemoryBank works with ChatGPT, ChatGLM, and BELLE, supports bilingual dialogs, and functions with or without the Ebbinghaus-based forgetting mechanism.
RESULTS
Retrieval Acc.
0.763
-0.051 vs SiliconFriend BELLE on English
Correctness
0.716
+0.237 over SiliconFriend ChatGLM on English
Coherence
0.912
+0.232 over SiliconFriend ChatGLM on English
Ranking
0.818
+0.301 over SiliconFriend ChatGLM on English
On a 10-day, 15-user simulated dialog benchmark with 194 probing questions, MemoryBank-powered SiliconFriend ChatGPT is evaluated for retrieval accuracy, response correctness, coherence, and human ranking. These results show that MemoryBank enables SiliconFriend ChatGPT to give more correct and coherent long-term responses than SiliconFriend ChatGLM and SiliconFriend BELLE despite slightly lower retrieval accuracy.
BENCHMARK
On a 10-day, 15-user simulated dialog benchmark with 194 probing questions, MemoryBank-powered SiliconFriend ChatGPT is evaluated for retrieval accuracy, response correctness, coherence, and human ranking. These results show that MemoryBank enables SiliconFriend ChatGPT to give more correct and coherent long-term responses than SiliconFriend ChatGLM and SiliconFriend BELLE despite slightly lower retrieval accuracy.
BENCHMARK
Correctness score on English probing questions for three MemoryBank-powered SiliconFriend variants.
KEY INSIGHT
MemoryBank with SiliconFriend ChatGPT reaches 0.716 correctness and 0.912 coherence even though its retrieval accuracy (0.763) is lower than BELLE and ChatGLM.
This is surprising because we usually expect higher retrieval accuracy to directly yield better answers, but MemoryBank shows that how memories are used can outweigh raw retrieval hits.
WHY IT MATTERS
MemoryBank unlocks persistent, personality-aware AI companions that can recall specific books, algorithms, and emotional history across at least 10 days of interaction.
Builders can now bolt MemoryBank onto diverse LLMs to get human-like forgetting and reinforcement, enabling long-term counseling, coaching, and secretarial agents without retraining base models.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…