Search

Agent memory

Agent memory works around language models’ lack of persistent memory by storing information externally and retrieving relevant entries into a task’s context. The document describes memory workflows, file-based and managed systems, consumer assistant features, and earlier research prototypes.

GPU Brain — Instant Agent Memory

GPU Brain offers a unified memory system for agents, enabling fast search, read, write, and linking of notes and context via markdown across hosted/local MCP, CLI, and REST interfaces, so agents recall the memory before answering.

[2610.34073] Memory Tax: Dissecting the Net Value of Agent Memory

The work evaluates whether adding memory to long-horizon agents provides net value, finding that many memory systems perform worse than simple retrieval-augmented generation; it introduces the concepts of memory benefit and memory tax to analyze gains and losses from memory design, and advocates evaluating memory by net value rather than raw performance, with open-source code to follow.

Bedrock AgentCore Long-Term Memory

The article explains adding a long-term memory layer to Bedrock AgentCore agents, combining in-session memory with a persistent memory that uses semantic search. It details the memory resource setup, code changes, and a Streamlit UI to manage and replay conversations.

The Memory Race Behind AI Computing | TechiesJournal

The document explains how memory types (DRAM and HBM) affect AI computing, covering capacity, bandwidth, and cost. It details HBM4's higher interface width and speeds, and notes ongoing competition for memory supply driven by AI server demand.

Discrete sortie of Hy-Memory-30B-A3B: Tencent's memory writer

Translated title. Original: Sortie discrète de Hy-Memory-30B-A3B : le rédacteur de mémoire de Tencent

The document profiles Hy-Memory-30B-A3B, a 30B mixture-of-experts model designed to write and manage persistent memory for long-duration agents, detailing its release, architecture, and evaluation setup; it discusses how memory-writing incurs cost and how the model is trained to operate across multiple frameworks, with references to competing variants and governance of benchmarks and incidents.

Groq Looks Up the Z-Axis: SRAM-First Inference Has Hit Its Capacity Wall

The piece discusses memory capacity for Groq's LPU-driven SRAM-first inference, noting SRAM and DDR5 footprints, questions about DRAM proximity to compute, and potential future 3D-stacked DRAM approaches from Groq and Cerebras, with limited official detail.

The Machine (computer architecture) - Wikipedia

The Machine is an HP/HPE research project to create a memory-centric server architecture using byte-addressable NVRAM, interconnected via photonics, with Linux++ and Carbon OS, targeting workloads like in-memory databases and real-time analytics; claimed speedups and a memory management model using a Librarian to segment memory into structured units.

TencentDB Agent Memory: Sharing Lessons Across AI Sessions

The article explains how TencentDB Agent Memory stores, organizes, and controls access to team knowledge across AI sessions. It describes the system’s components and reported PersonaMem results, while noting that the benchmark comes from the project README.

[2610.27377] REM: What to Store, When to Recall, What to Forget in Experience Memory for Coding Agents

REM is a memory framework for autonomous code- writing agents that stores, recalls, and forgets experiences within the think-act loop. It improves end-task performance and reduces cost by using phase-structured, compact summaries prepared at task start and gated recalls during execution, with eviction after one exposure.