Semantic-TVM: Trustworthy Virtual Memory for Memory-Augmented AI Agents

A new arXiv paper proposes Semantic-TVM, a virtual memory design that keeps sensitive values protected while still letting agent workflows run on remote language models. The approach targets memory-augmented and tool-using agents, where retrieved memories, tool calls, and intermediate observations can leak private data. It aims to move past one-way masking, which hides values but also blocks the trusted execution they are needed for.

CoMem Paper Proposes Shared and Individual Memory Design for LLM Multi-Agent Systems

A new arXiv preprint introduces CoMem, a memory framework for LLM-driven multi-agent systems that combines collective knowledge with agent-specific memory. The authors argue that most existing approaches rely on flat, unstructured memory, which limits how agents learn and improve over time. The work targets better long-term cooperation and performance in evolutionary multi-agent setups.

Pull Router Cuts Token Costs in Long LLM Conversations via Lazy Memory Loading

A new arXiv paper proposes Pull, a session router that keeps conversation history in an addressable store instead of injecting all prior turns into the context window. Rather than summarizing or truncating old dialogue, the system materializes only the historical state needed at each step, aiming to avoid the quadratic token growth of full-context prompting without permanently losing information. The approach targets multi-hundred-turn, stateful LLM sessions.

arXiv Paper Describes Persistent Memory Poisoning Attack on Harness-Based LLM Agents

A new arXiv preprint examines how harness-based LLM agents, which combine memory, tool use, and runtime control, can be compromised through stored malicious instructions. The authors argue that once such instructions enter an agent's persistent memory, they can continue to influence later behavior, creating security and privacy exposure. The work frames memory poisoning as a distinct risk for agent architectures that retain context across sessions.

MemRiskBench Benchmark Targets Memory Risks in Long-Horizon LLM Agents

A new arXiv paper introduces MemRiskBench, an evaluation framework for long-horizon LLM agents that accumulate memory across sessions. It measures per-risk failure rates for issues such as stale facts, conflicting updates, cross-user data leakage, reuse of revoked memories, and decay of constraints, which aggregate scores tend to obscure. The work argues for trace-aware evaluation that preserves these distinct risk categories rather than collapsing them into a single number.

Paper Proposes Self-Evolving Memory for Generative Recommendation

A new arXiv paper addresses how generative recommender systems can keep up with user preferences that shift over time. The authors propose a self-evolving memory mechanism intended to let these models update their behavior as tastes change. The work is categorized as a cross-list submission under arXiv cs.AI.

Study Ties Data and Memory Scaling to a Single Predictive Spectrum

A new arXiv paper argues that the benefit a model gains from more data and the amount of learned memory it needs are both determined by one predictive-energy spectrum in a positive-entropy autoregressive retrieval source. In this framework, each coordinate's contribution is the product of its query probability and a per-coordinate term, linking the two resources analytically. The work offers a theoretical account of how data volume and memory capacity trade off in autoregressive prediction.

LIMBO: Inference-Time Memory and Budget Optimization for Lifelong LLM Agents

A new arXiv paper introduces LIMBO, a method that optimizes memory use and compute budgets at inference time so LLM agents can keep learning new tasks without losing earlier skills. It builds on experience replay, which feeds past interactions back into the agent, and targets the cost and context limits that arise as agents run in long, evolving workflows. The work is positioned as a general approach to lifelong capability retention for deployed agents.

MMLA: Memory-Mediated Learning Architecture for Predictive Dual-State Adaptation

A new arXiv paper proposes a Memory-Mediated Learning Architecture (MMLA) that splits a learning system into slow base parameters, a bounded numerical policy carrier, and a bounded authoritative memory store. Its Predictive Dual-State Adaptation mechanism uses feedback to update the policy carrier while handling a problem-state component. The work is presented as a preprint on arXiv under cs.LG.

Paper proposes transport criterion for exact record omission in delta attention

A new arXiv paper examines the gap between a model appearing to forget a deleted record and its recurrent state still retaining that record's influence. The authors store the state difference produced right after a record is written and track it through subsequent processing steps, deriving a condition under which the omission is exact along with its computational cost and a replay-based certificate. The work targets attention mechanisms with recurrent memory, where deleting data from context does not necessarily remove its effect from internal state.

BudgetBench: Budget-Tiered Protocol for Evaluating Memory in Local LLM Agents

Researchers introduce BudgetBench, a protocol and pilot test harness for assessing memory strategies in locally run large language model agents. The work treats active context as a limited resource, accounting for memory capacity, prefill latency, cache growth, and service targets when deciding how many input tokens a call can afford. It aims to give a structured, budget-aware way to compare agent memory approaches under constrained hardware and latency conditions.

FluctlightDB Proposes Cue-Driven Memory Model for AI Agent Data

A new arXiv paper introduces FluctlightDB, a data model intended for AI agent memory rather than relational or vector querying. It argues that existing systems were designed around predicate matching or nearest-neighbor search, and neither fits recall triggered by cues and weighted by provenance. The work positions agent memory as a distinct storage problem needing its own abstractions.

Paper Proposes Auditable Method for Verifying Deletion of Stored Facts in AI Models

The paper installs a support-vector gate inside a frozen Gemma 3 model and records which stored keys and values correspond to each exchange, allowing a deletion request to be traced to specific memory entries. It argues that an assistant can stop repeating a fact without actually removing it, so the work separates genuine deletion from mere output suppression. The authors present this as a step toward deletion requests that can be audited rather than taken on trust.

New Preprint Introduces ConvMem, a Convolutional Memory Method for Long-Context Reasoning

An arXiv preprint proposes ConvMem, a convolutional memory technique aimed at helping large language models reason over documents that exceed their fixed context windows. The approach builds on prior sequential strategies such as MemAgent, which extend usable context by processing text in stages. The paper was posted to the AI and computational linguistics categories on arXiv.

Study examines procedural memory reuse and interference in language web agents

A new arXiv paper investigates what happens to language agents when the routines they have memorized no longer fit their environment. Combining a retrospective, human-assisted analysis with controlled web-task experiments, the authors test when stored procedures can still be successfully reused and when they interfere with one another. The work addresses a core assumption behind procedural memory in autonomous agents.

Study separates what language agents store in memory from what they use to answer

An arXiv paper examines memory management for persistent language agents, arguing that the experiences an agent keeps in storage should be treated separately from the evidence it retrieves for any individual question. The authors note that a superseded fact can mislead an answer about the current state while still being essential for queries about the past. The work frames forgetting as a deliberate design choice in agent architectures.

EviMem proposes evidence-gap-driven iterative retrieval for long-term conversational memory

Researchers present EviMem, a retrieval method for long-term conversational memory that identifies gaps in the evidence gathered so far and iteratively fetches additional material across past sessions. The approach targets temporal and multi-hop questions where a single retrieval pass typically fails to locate relevant information. The paper is available as a revised version (v2) on arXiv.

Fortunate Recall: Ontology-Driven Memory Lifecycle Management for LLMs

A newly posted arXiv paper tackles the problem of AI assistants accumulating unstructured personal information until their long-term memory becomes cluttered and less reliable. The proposed method organizes stored facts using an ontology and applies lifecycle rules that determine which memories should be retained, updated, or discarded over time. The goal is to keep conversational agents internally consistent during extended interactions while improving the precision of memory retrieval.

PRAGMA Benchmark Evaluates Personalized Guidance with Memory Alignment in Lifelong Conversations

A new arXiv paper introduces PRAGMA, a benchmark for measuring how well large language models align stored user memory with the advice they deliver in long-running conversations. The work targets a key weakness of personalized assistants: as dialogue histories grow, working from complete logs becomes inefficient and error-prone. PRAGMA offers a standardized way to evaluate memory use in lifelong conversational systems.

ROAM Framework Organizes Atomic Memories for Long-Term LLM Agents via Semantic Relations

Researchers have introduced ROAM, a method for structuring the external memory of language-model agents as fine-grained atomic units linked by semantic relations. The approach is designed to improve how agents retrieve and compare stored observations across extended interactions. The work appears in a newly published preprint on the arXiv repository.

arXiv Paper Proposes Memory Compression for High-Fanout Agent Sandboxes

A new arXiv preprint describes a memory compression approach for agent workloads in which one task launches many concurrent sandbox sessions. Because those sessions are spun up from a common template and run similar trajectories, the paper argues they share exploitable structure that can reduce the memory footprint. The method is aimed at easing the memory bottleneck that appears as agent fanout grows.

Paper Proposes Environment-Probing Curation for Enterprise Agent Memory

A new arXiv paper addresses how persistent memory is being adopted in production agent platforms so long-horizon agents can retain experience across sessions. The authors argue that curator agents limited to completed task traces can lock in mistakes and draw overly broad conclusions from incomplete evidence, and propose probing the environment as an alternative approach to memory curation.

Agent-Side Memory as Guidance for Steerable Action Models in Long-Horizon Manipulation

A new arXiv paper argues that long-horizon robot manipulation needs memory, but that this memory does not have to live inside the action policy itself. The work proposes placing memory on the agent side and using it as guidance to steer action models, rather than relying solely on vision-language-action policies paired with planners and geometric tools. The approach targets tasks that span many steps, where extra depth or calibrated geometry is sometimes added to the control stack.

Kepler Compute pitches ferroelectric memory as fast, cheap alternative to HBM

US startup Kepler Compute says it can address the RAM supply crunch with ferroelectric memory technology positioned as a replacement for HBM. The company claims the approach could scale manufacturing capacity for memory quickly and at lower cost, with Globalfoundries mentioned in coverage of the effort. Details on timelines, production volumes and customer commitments were not provided.

Kepler Computing Says New Chip Design and Material Could Ease Memory Supply Bottleneck

A previously little-known startup, Kepler Computing, says it has developed a chip design approach along with a proprietary material that could relieve the memory supply constraints behind recent price spikes. The company has not yet disclosed full technical details or independent verification of its claims. If the approach works at scale, it could reduce pressure on a market where memory costs have climbed sharply.

arXiv Paper Examines How Memory Structure Is Evaluated in LLM Agents

A newly revised arXiv preprint looks at long-term memory frameworks used by LLM-based agents and chat assistants, which store reusable knowledge, recall user preferences, and support reasoning. The authors argue that as memory architectures grow more complex, assessing them becomes progressively harder. The work focuses on methods for evaluating how memory is structured in these systems.

Hugging Face shows how to give coding agents persistent memory you control

A new Hugging Face blog post walks through building a memory layer for coding agents that remains under the developer's ownership instead of depending on external services. It explains how to store and reuse context such as past tasks, decisions, and code history so an agent can resume work across sessions. The guide favors open, self-hostable setups for keeping project knowledge local.

WHY IT MATTERS ↘Memory is emerging as a key differentiator for coding agents, but most commercial memory features are vendor-hosted, locking teams' accumulated project knowledge into external services. An open, self-hostable alternative addresses governance, data-residency, and lock-in concerns, making it more viable for enterprises to adopt agents while keeping proprietary context in-house.