Daily Brief — Oct 16, 2026: Agent Memory

🛰 AI Brief — Oct 16, 2026

How to read prio and sources

prio N is the radar’s practical-relevance score for this item (higher runs first; items at or below the noise threshold are filtered out as noise). Under each signal: Concepts / Entities are graph links; Source / N sources list every outbound link for that story.

🥇 STRATA: Self-Maintaining Structured Experience Graphs for Self-Evolving Language Agents · prio 12

For builders creating autonomous agents, this paper addresses the weak area of agent memory by identifying failure modes in naive single-level vector retrieval, such as catastrophic negative transfer and uncontrolled context bloating. It offers concrete architectural patterns—specifically layered abstraction hierarchies and contrast edges—to maintain persistent agent experience efficiently over time. Concepts: Agent Memory Agents Context Engineering Source: arxiv.org

🥈 MemCode: Discrete Semantic Representations for Long-Term Agent Memory · prio 10

For builders constructing multi-session AI agents, this paper addresses the common failure modes of continuous embedding retrieval by introducing discrete semantic quantization and topological co-occurrence retrieval. It directly informs the community’s weak understanding of agent memory architectures while demonstrating significant speed improvements over prior memory systems like LightMem and MemoryOS. Concepts: Agent Memory Agents Embeddings Entities: gpt-4o-mini Qwen3-30B-A3B-Instruct-2507 Source: arxiv.org

🥉 EvoWeave: Composing Reusable Agent Experience through Operation Level Graphs · prio 10

The paper directly addresses agent memory architecture, a weak area for the community, by proposing a graph-based approach to reuse abstract operations across agent workflows. Evaluation on AppWorld benchmarks shows improvements of 6-7 percentage points, demonstrating measurable gains from operation-level memory composition for multi-step agents. Concepts: Agent Memory Agents Source: arxiv.org

4️⃣ Multimodal Agent Memory: A Framework for Formation, Representation, and Adaptive Management · prio 9

Agent memory is an identified weak area for the community, and this paper directly addresses it with a structured framework distinguishing memory formation from three tiers of management (static, dynamic, self-evolving). For builders developing long-horizon agents, understanding these design patterns and the tradeoffs between representation fidelity, temporal validity, and policy adaptation is essential to move beyond fixed memory architectures. Concepts: Agent Memory Source: arxiv.org

5️⃣ Context-Guided Model Routing for Stateful Multi-Turn Agents · prio 8

For builders running multi-turn AI agents, naive per-turn model routing breaks prompt caching and degrades performance if session history is ignored. Factoring in asymmetric cache costs and session context allows developers to offload turns to cheaper models without sacrificing agent accuracy. Concepts: Agents Context Engineering Tool Use Source: arxiv.org

Knowledge Gaps

Topics the AI stream keeps raising that the knowledge base hasn’t sufficiently covered yet — candidates for what to learn next. Agent Memory · Embeddings

🧪 Research Papers (11)

• prio 8 When Is Smaller Greener? Right-Sizing Language Models for Agentic AI Concepts: Agents LLM Evals Open Source LLMs Source: arxiv.org
• prio 8 Embedding Geometry Shapes LLM Routing: Tracing Failure Modes and Improving Generalization Concepts: Embeddings Entities: Qwen2.5-7B-Instruct Source: arxiv.org
• prio 8 ListQA: A Benchmark for Evaluating List-Formatted Factual Knowledge Retrieval in Large Language Models Concepts: LLM Evals RAG Source: arxiv.org
• prio 8 AgentCIScope: Grounded Contextual Integrity Evaluation for LLM Agents Concepts: Agents LLM Evals Source: arxiv.org
• prio 8 Plant Memory: Knowledge Extraction and Verification System for Agent-Accessible Industrial Maintenance Ontologies Concepts: Agent Memory Agents MCP Entities: SAP Source: arxiv.org
• prio 8 Plan or Retrieve First? A Scale Study of Tool Discovery Architectures Concepts: Tool Use Agents Context Engineering Entities: Google Gemini 2.5 Flash GPT-OSS 120B Source: arxiv.org
• prio 7 Evaluating Agentic Capabilities Before Post-Training: An End-to-End Approach Concepts: Agents LLM Evals Source: arxiv.org
• prio 7 DeLoop: Repeated Actions as Self-Supervision for Long-Horizon Agents Concepts: Agents Context Engineering Code Agents Entities: Qwen LLaMA Gemini Qwen3-8B Source: arxiv.org
• prio 7 Study Finds Static Skill Inconsistencies Do Not Directly Correlate to Runtime Risks in LLM Agents Concepts: Agents Tool Use LLM Evals Source: arxiv.org
• prio 6 EvoProbe: Evaluation-Driven Benchmark Evolution for LLM Agents Concepts: Agents LLM Evals Source: arxiv.org
• prio 6 Verification Turns Capability Gaps into Routing Decisions: Small Language Models for Clinical Documentation Concepts: Agents LLM Evals Source: arxiv.org