AuthorsLei Liu, Xiaoyan Yang, Yue Shen et al.
TL;DR
Think-in-Memory (TiM) stores and retrieves LLM-generated inductive thoughts via LSH-based hashing, boosting contextual coherence on Chinese GVD from 0.428 to 0.665 over SiliconFriend (+0.237).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Memory-augmented LLMs rely on iterative recalling and reasoning, which easily produce biased thoughts, i.e., inconsistent reasoning results over the same history.
In long-term tasks like medical AI assistants, this repeated reasoning over raw text can miss earlier symptoms and reduce diagnosis accuracy in multi-turn conversations.
HOW IT WORKS
Think-in-Memory (TiM) combines an Agent A, a hash-based Memory Cache M, a Hash-based Mapping F(·), and Insert Forget Merge organization operations over inductive thoughts.
You can view TiM like a brain that stores distilled thoughts instead of raw events, using LSH buckets as shelves and Merge as compressing similar memories.
By recalling stored inductive thoughts instead of re-parsing full histories, TiM enables consistent reasoning and efficient long-term retrieval beyond a plain context window.
DIAGRAM
This diagram shows how TiM performs Stage-1 Recall and Generation and Stage-2 Post-think and Update for each new query.
DIAGRAM
This diagram shows how TiM is combined with ChatGLM and Baichuan2 and evaluated on GVD, KdConv, and RMD with three human-judged metrics.
PROCESS
Stage-1 Recall and Generation
Given a new query Q, Think-in-Memory uses Agent A, Hash-based Mapping F(·), and Memory Cache M to recall top-k relevant inductive thoughts and generate a response.
Stage-2 Post-think and Update
After producing the response R, Think-in-Memory lets Agent A post-think on the Q-R pair and generate new inductive thoughts describing entities and relations.
Insert operation
Think-in-Memory applies the Insert operation by hashing each new inductive thought with F(·) and storing it in the corresponding group inside Memory Cache M.
Forget and Merge operations
Think-in-Memory periodically uses Forget to remove counterfactual thoughts and Merge to combine similar thoughts with the same entity within each hash group.
KEY CONTRIBUTIONS
Think-in-Memory long-term mechanism
Think-in-Memory introduces an LLM-agnostic Memory Cache M of inductive thoughts with Hash-based Mapping F(·), enabling recall without repeated reasoning over raw Q-R histories.
Insert Forget Merge organization
Think-in-Memory formalizes Insert, Forget, and Merge operations over thoughts within hash groups, mirroring human cognitive processes for dynamic memory evolution.
Hash-based retrieval for long-term conversations
Think-in-Memory integrates Locality-Sensitive Hashing into F(·), reducing retrieval time from 0.6287 ms to 0.5305 ms and improving contextual coherence on GVD and KdConv.
RESULTS
Contextual Coherence (GVD Chinese)
0.665
+0.237 over SiliconFriend
Response Correctness (GVD Chinese)
0.605
+0.187 over SiliconFriend
Retrieval Accuracy (KdConv Film ChatGLM)
0.920
top-5 recall within TiM memory cache
Retrieval Time per query
0.5305 ms
0.0982 ms faster than baseline pairwise similarity
On the Chinese part of GVD, which tests long-term open-domain dialogue, Think-in-Memory with ChatGLM raises contextual coherence from 0.428 to 0.665 and response correctness from 0.418 to 0.605 over SiliconFriend. On KdConv Film with ChatGLM, Think-in-Memory achieves 0.920 retrieval accuracy and improves response correctness from 0.657 to 0.827.
BENCHMARK
On the Chinese part of GVD, which tests long-term open-domain dialogue, Think-in-Memory with ChatGLM raises contextual coherence from 0.428 to 0.665 and response correctness from 0.418 to 0.605 over SiliconFriend. On KdConv Film with ChatGLM, Think-in-Memory achieves 0.920 retrieval accuracy and improves response correctness from 0.657 to 0.827.
BENCHMARK
Contextual Coherence on the Chinese part of the Generated Virtual Dataset (GVD).
KEY INSIGHT
Despite using only distilled inductive thoughts, Think-in-Memory reaches 0.920 retrieval accuracy on KdConv Film with ChatGLM using top-5 recall.
This is surprising because many expect discarding raw dialogue text to hurt retrieval, yet structured thoughts plus LSH actually improve both efficiency and accuracy.
WHY IT MATTERS
Think-in-Memory unlocks LLM agents that maintain consistent reasoning paths over long-term conversations by recalling inductive thoughts instead of re-reading entire histories.
Builders can now bolt TiM onto existing LLMs like ChatGLM or Baichuan2 to get efficient, evolvable long-term memory without modifying model architectures.
Guilin Zhang, Wei Jiang et al.
· 2026
A-MAC scores candidate memories using Utility, Confidence, Novelty, Recency, and Type Prior combined by a learned linear admission policy with Algorithm 1 A-MAC Memory Admission. On the LoCoMo benchmark, A-MAC achieves F1 0.583 and 2644 ms latency, improving F1 by 0.042 and reducing latency by 1187 ms compared to A-mem.
Robbyant Team, Zelin Gao et al.
arXiv 2026 · 2026
LingBot-World combines a Data Engine, Fundamental World Model, Action-Conditioned World Model, and Post-Training causal adaptation to turn a 28B-parameter video generator into a real-time interactive world simulator. On the VBench benchmark, LingBot-World achieves a dynamic degree of 0.8857 versus 0.7612 for Yume-1.5, while also improving imaging quality to 0.6683.
Manoj Madushanka Perera, Adnan Mahmood et al.
· 2026
AgenticAI-DialogGen chains ChatPreprocessor, KnowledgeExtractor, TopicAnalyzer, KnowledgeGraphBuilder, PersonaGenerator, DuelingChat Agent, ConversationValidator, ConversationRefiner, QAGeneration, and PostProcessing to turn raw multi-session chats into topic-guided, persona-grounded conversations with explicit short- and long-term memories. On the TGC / KG memory QA benchmark, Mistral-7B fine-tuned within AgenticAI-DialogGen achieves 87.36 F1, compared to GPT-4’s 83.77 F1 in a zero-shot setting on the same task.
Answers use this explainer on Memory Papers.
Checking…