Tencent Yuanbao and King's College London's NLP team jointly released GroupAssistBench, built from 50 real group chats and 80,299 historical messages into 2,395 evaluation tasks, splitting agent memory into L1 factual recall and L2 open-ended service. Across 14 configurations, stronger recall did not imply better service: EverMemOS scored 0.621 on L1 versus 0.570 for full context, yet only 0.247 on L2 versus 0.431. The paper, code and dataset are all public.
Agent Memory
智能体记忆
Mechanisms for agents to store and retrieve historical information across turns, covering writing, retrieval and effective use i
Curated Intelligence
Amazon Bedrock Managed Knowledge Base introduces Agentic Retrieve, adding user memory on top of document retrieval. The feature extracts durable facts from past conversations, selects those relevant to the current question, and injects them into the retrieval request with per-user and per-document isolation. It targets support assistants, in-product document chat, and internal policy assistants.
Anthropic studied how Claude retains knowledge across sessions and found that selectively recording strategies and rules outperforms exhaustive logging, which floods models with outdated data. Its post-session "Dreaming" process analyzes past transcripts during idle periods to correct faulty notes and build structured knowledge collections.
Researchers from nine top universities have jointly released a 126-page roadmap for always-on agents, framing the shift in personal AI from single-model features toward coordinated perception, memory and action architectures. The report directly tackles what counts as truly always-on and what an agent should or should not do when the user is absent, offering the field a first systematic architectural reference.
A joint MIT and Meta paper introduces ReHarness, which revises an agent's previously retrieved notes using the outcome of each task instead of only appending new ones, preventing stale or over-broad lessons from misleading the agent. Across three benchmarks and three frontier models it averaged 0.240 normalized gain versus 0.168 for the best raw-trajectory baseline, while using roughly one-third the tokens of replaying full raw history on CL-Bench with Claude.
The article argues that the bottleneck in agentic AI has shifted from retrieval to context governance, proposing a split of context, memory, and knowledge into three distinct components orchestrated by a 'context as control plane' paradigm. Its core claim: the model generates, while a surrounding context engine decides what the model can see, suppressing hallucination and cost blowup at the architectural level.
Agent memory today is scattered across AGENTS.md files, /memories directories, Copilot's repo-side facts and MCP knowledge graphs, none of which can read the others, so the same lesson gets rebuilt repeatedly. The author argues the real gap is not another tool but a shared format and convention akin to OpenTelemetry, using LoreKit's plain-Markdown offline store as a concrete, inspectable example.
Anthropic's engineering team detailed a four-layer memory architecture for Claude spanning raw session logs, self-written notes, structured wikis, and auto-injected context, plus a post-session 'dreaming' step that revisits transcripts to fix and reorganize notes. Pokemon and database-QA tests show memory quality, not volume, drives performance, with Opus 5 the first model to form a complete learning loop.
Anthropic's engineering blog uses human memory as an analogy to break down the components of AI agent memory, framing it under agents and context engineering. It is a conceptual engineering write-up, with the middle of the text omitted and no verifiable methods or parameters provided.
Anthropic published research titled 'The components of agent memory,' splitting Claude agent memory into four layers: session logs, memory notes, knowledge representations and context injection, plus a 'dreaming' mechanism that lets the model review logs, correct errors and distill rules during idle time. On the CL-Bench database QA task, memory-enabled Claude Opus 5.5 and Sonnet 5.5 scored 0.838 and 0.833 on average, well above their no-memory baselines.