All AI memory papers

Episodic Memory is the Missing Piece for Long-Term LLM Agents proposes an architecture where in-context memory, external memory, and parametric memory are coordinated via consolidation, encoding, and retrieval to realize long-term, instance-specific, contextual episodic traces. Episodic Memory is the Missing Piece for Long-Term LLM Agents contributes a five-property taxonomy, a three-way memory categorization (in-context, external, parametric), and a roadmap of six research questions instead of benchmark gains.

PREMem builds long term dialogue memory by combining Episodic Memory Extraction, Pre Storage Memory Reasoning, semantic clustering, a persistent memory pool, and an inference phase over enriched memory fragments. PREMem reaches 71.4 LLM as a judge on LongMemEval with gpt 4.1 base, a +15.5 gain over HippoRAG 2 and +9.6 over A Mem.

ReMe manages procedural memory through experience acquisition, experience reuse, and experience refinement, combining multi-faceted distillation, context-adaptive reuse, and utility-based deletion into a single lifecycle. On BFCL-V3 and AppWorld, Qwen3-8B with ReMe (dynamic) achieves 34.94% Avg@4 vs 27.65% for the No Memory baseline, and 55.03% Pass@4 vs 46.20%, showing that self-evolving memory can substitute for model scale.

RGMem builds a multi-scale memory state using Microscopic Evidence Space DL0, Structured Knowledge Space G, and renormalization operators RK1, RK2, RK3 to evolve user profiles. On PersonaMem with GPT-4.1, RGMem reaches 74.01% Avg., beating Memory OS by 8.98 points.

Semantic Anchoring enriches conversational memory by combining a hybrid memory store with dense and symbolic indexes, structured memory representation tuples, hybrid storage and indexing, and a retrieval scoring method. On MultiWOZ-Long, Semantic Anchoring reaches 83.5% Factual Recall and 80.8% Discourse Coherence, beating Entity-RAG by 7.6 and 8.6 points respectively.

SGMem organizes long conversations via SGMem Construction and Management, SGMem Usage, sentence level graphs, and multi hop retrieval over sessions, rounds, turns, summaries, facts, and insights. SGMem achieves 0.700 Accuracy (Top 5) on LongMemEval and 0.526 on LoCoMo, beating the RAG-SMFI baseline at 0.676 and 0.510 respectively.

StoryBench evaluates long-term memory by embedding LLMs in Dynamic Narrative and Multi-Turn Decision-Making, two Task Modes for Evaluating LTM, and Tailored Metrics for Assessing LTM Models over a 311-scene, 86-choice graph. StoryBench shows Doubao1.5-pro achieving 80.98% Overall Acc in Immediate Feedback mode on The Invisible Guardian while Claude 3.5 Sonnet attains the highest Success Count of 8, revealing distinct strengths and weaknesses.

TeleMem converts interactions into unified semantic nodes via the representation layer, organizes them in a memory graph with Insert and ReInsert, and reads them using closure-based retrieval and a ReAct-style multimodal agent. On ZH-4O, TeleMem reaches 86.33% QA Accuracy, beating the Mem0 baseline at 70.20% and the RAG baseline at 62.45%.

Test-time regression uses memorization as regression, memory retrieval, and test-time regression layers to reinterpret sequence architectures as solving a regression problem over key value pairs during the forward pass. This unification shows how linear attention, state space models, fast weight programmers, online learning layers, and softmax attention are all instances of the same framework and explains phenomena like linear attention’s failures and the role of query key normalization.

TokMem adds a Memory Bank of trainable memory tokens to a frozen Transformer backbone, using memory routing, conditional generation, and renormalization to store and recall procedures. On Super-Natural Instructions and APIGen function-calling, TokMem reaches 67.0 ROUGE-L and 99.1 tool-selection F1, surpassing Replay Memory and LoRA fine-tuning with far fewer parameters.

TurboQuant combines MSE Optimal TurboQuant, Inner-product Optimal TurboQuant, QJL, Random Rotation Matrix Π, and Lloyd-Max Quantizer to quantize vectors online with near-optimal distortion-rate guarantees. TurboQuant matches the Shannon lower bound within a factor of √(3π/2)≈2.7 for MSE and achieves absolute quality neutrality for KV cache quantization at 3.5 bits per channel compared to full-precision baselines.

PI-LLM uses a synthetic key–value retrieval task, Interference Endurance Score (IES), per-key forget prompts, and a mock QA reset to stress-test working-memory-like behavior under proactive interference. PI-LLM finds a universal log-linear decay in retrieval accuracy across 0.6B–637B-parameter LLMs as interference grows, revealing that parameter size, not context window length, predicts interference robustness.

Understanding Transformer from the Perspective of Associative Memory reframes Softmax Attention, Linear Attention, FFN, and DeltaNet as instances of a unified associative memory with explicit memory capacity and update rules. Using this lens, Understanding Transformer from the Perspective of Associative Memory derives retrieval SNR for different kernels, unifies attention and FFNs, and proves that DeltaFormer achieves circuit complexity beyond TC0, reaching NC1 expressivity.

Understanding Users' Privacy Perceptions Towards LLM's RAG-based Memory analyzes users' mental models, privacy calculus, and expectations around RAG-based memory across generation, management, usage, and updating. Understanding Users' Privacy Perceptions Towards LLM's RAG-based Memory finds users demand explicit consent, fine-grained editing and deletion, and visibility into inferred information to trust RAG-based memory systems.

MEXTRA crafts black box attacking prompts and automated diverse prompt generators that target the memory module, similarity scoring function, retrieval depth, memory size, and LLM backbone. MEXTRA extracts 50 queries from a 200 record EHRAgent memory and 26 from RAP, with extracted efficiency up to 0.42 compared to weaker baselines without workflow aligned prompts.

WebATLAS combines a Planner, Actor, Critic, and Multi-layered Memory (Working Memory, Cognitive Map, Semantic Memory) to simulate and score actions before executing them on the web. On WebArena-Lite, WebATLAS achieves 63.0% average success versus 53.9% for Plan-and-Act, a +9.1 point gain without website-specific fine-tuning.

WorldMM dynamically coordinates Episodic Memory, Semantic Memory, Visual Memory, an Adaptive Retrieval Agent, and a Response Agent to answer queries over hour- to week-long videos. On five long video QA benchmarks, WorldMM-GPT reaches 69.5% average accuracy, beating M3-Agent’s 55.1% by 14.4 points and the best prior memory baseline HippoRAG’s 57.0% by 12.5 points.

Zep builds memory using Episode Subgraph, Semantic Entity Subgraph, Community Subgraph, and a three-stage Search–Reranker–Constructor pipeline over the Graphiti temporal knowledge graph. On LongMemEval, Zep with gpt-4o scores 71.2% vs a 60.2% full-context baseline and reduces average latency from 28.9 s to 2.58 s.

AriGraph builds a joint world model by combining semantic memory, episodic memory, semantic search, episodic search, and the Ariadne cognitive architecture into a single evolving graph. On NetHack, AriGraph lets Ariadne reach a score of 593.00 with room-only observations, compared to 341.67 for NetPlay with the same input and 675.33 for NetPlay with full level observations.

CAMELoT augments frozen LLaMA2‑7B with Read, Augment, and Write associative memory modules that maintain consolidated key–value slots with novelty and recency‑aware updates. On Arxiv long‑context language modeling, CAMELoT reaches 3.60 average perplexity versus 5.12 for LLaMa2‑7B, a 29.7% reduction without any retraining.

EMG-RAG combines Editable Memory Graphs, a two-stage MDP for Selecting Memories on EMGs, and Data Collection from real smartphone assistants to support editable, graph-structured personal memories. On a 2,500-user business dataset, EMG-RAG with GPT-4 reaches 75.99 BLEU on question answering, a +11.83 gain over M-RAG.

Deciphering the Interplay of Parametric and Non-parametric Memory instruments causal mediation analysis, Experiment 1, Experiment 2, and Path Specific Effects (PSE) inside ATLAS to trace how parametric and non-parametric memories compete token-by-token. Deciphering the Interplay of Parametric and Non-parametric Memory reports a strong shift toward counterfactual answers in altered contexts, with a t-test p-value of 1.60e-4 and Cohen’s d of -0.9851 for non-parametric versus parametric behavior.

Do LLMs dream of elephants studies how a self-attention layer, value matrix, embedding matrix, latent concept association task, and context hijacking prompts interact to implement associative memory in transformers. Do LLMs dream of elephants proves theoretically (Theorem 1, Theorem 4) and empirically that a one-layer transformer can achieve arbitrarily small error on latent concept association by using the value matrix as associative memory.