retrieval-augmented generation、向量檢索、知識庫、document parsing
不同使用者面對相同任務，agent 應該有不同做法——這篇把「個人化」從 prompt 層面深入到 RL 訓練層面，提出三組件框架（PARPO + 偏好解糾纏獎勵模型 + PSGM 圖記憶）讓 agent 在學習時就把個體差異嵌入。
arxiv: 2605.23382你以為 agent 的效能瓶頸是「prompt 太長」——錯了。真正的瓶頸是 token 生成速度與 KV-cache 管理，serving 系統需要為此重新設計。
arxiv: 2605.26297在 agent loop 裡找資料，直接用 grep（字串比對）往往比向量搜尋（semantic search）更準——但真正影響結果的是你用哪個 harness，不只是搜尋方法本身。
arxiv: 2605.15184提出 memorywire，一個讓 mem0、Letta、Cognee、pgvector 等記憶框架「插頭共用」的 JSON 電線格式，定義 5 個記憶操作與 4 種記憶類型，附帶 5 個後端 adapter 的開源實作。
arxiv: 2606.01138不再把過去成功/失敗經驗轉成文字塞進 context，改成在 LLM 隱層向量空間做檢索與整合，12/13 benchmark 最佳，零樣本跨域遷移提升 16.32%。
arxiv: 2606.01041Agent 自我演化不只是改 prompt 一件事；APEX 同時演化三層：執行環境（harness）、行為原則（principles）、工作流程拓撲（topology），在有 114 筆真實任務軌跡的生產 agent 上驗證。
arxiv: 2606.15363讓 agent 把完整的執行軌跡（怎麼一步步完成任務）存進共享庫，讓其他 agent 之後能撈出來參考，而不是每次遇到類似任務都重新摸索。
arxiv: 2606.19911大家評估 Agent 記憶體都只看「答對幾題」，但忘了問「這個設計貴不貴、知識更新後會不會崩掉」——這篇從資料庫角度把記憶體拆成四個可以分別量測的系統模組。
arxiv: 2606.24775把「用戶模型」從一堆文字筆記，升級成可以直接執行的 Python 程式碼，讓 Agent 不只能「查詢事實」，還能對用戶狀態做計算與邏輯推理。
arxiv: 2606.16707把記憶體存取從「工具調用（100ms）」變成「同程序存取（100μs）」，agent 在 12 步任務中的冗餘行為從 7.2 次降到 0——延遲不只是工程效能問題，它從根本上改變 agent 推理的品質。
arxiv: 2607.05690用四個設計維度重新打造電商網站，讓 AI 瀏覽器 agent 完成購物任務的成功率從 49% 提升到 89%。
arxiv: 2607.12056把 agent 系統拆成三張圖來描述，讓元件可以重複使用、執行路徑可以視覺化追蹤，降低複雜多 agent 系統的設計與除錯成本。
arxiv: 2607.13474用統一公式整理「agent 怎麼從自己的執行經驗中學習和進化」，把所有方法按「更新什麼」和「用什麼訊號更新」兩個維度分類。
arxiv: 2607.13104持久記憶讓 agent 記得過去，但也讓攻擊者能「先植入一段話，等之後特定情境出現時 agent 自動執行惡意行為」——這篇建立 1,227 個案例的 benchmark，量化這個威脅有多嚴重。
arxiv: 2607.14651把「記憶管理」當成獨立技能來訓練，完全不改模型的任務行為，光是優化記憶就讓 32B 開源模型在長程遊戲上達到 Claude Opus 4.5 的水準（Crafter 51.4% vs 49.5%）。
arxiv: 2607.01224RippleMem boosts LongMemEval-S accuracy by up to 11.87% via associative memory spreading while cutting graph construction cost to 1/30; Total Recall at What Cost? measures 18–69% prediction error in memory system serving costs with no system winning both cost and accuracy; MESA's dynamic structure selection achieves 8.5% higher accuracy on AMA-Bench while saving 41% of evidence tokens
Three papers examining AI Agent capabilities and limits from different angles: AutoMem shows memory management is a learnable skill — optimizing memory alone lifts a 32B open-source model to top commercial model levels; Shadow Evaluation tests whether frontier Agents can do open-ended AI research using real NeurIPS submissions — the answer is no, Agents can engineer but cannot research; Adaptive Adversaries reveals that existing safety benchmarks severely underestimate threats — adding adaptive multi-turn attackers jumps ASR from 0–1% to 14%. Together, these three papers deliver a sobering lesson: know where Agents can automatically improve, where they cannot, and that your security testing is probably insufficient.
Three papers tackling three core agent-platform challenges: MyAG introduces a graph-theoretic decomposition of agent systems into component / workflow / search layers; a self-improvement survey unifies the entire 'how agents evolve from experience' landscape under one formula; and MemPoison reveals persistent memory as the most vulnerable attack surface, with a 1,227-case benchmark. Together they cover: how to architect → how to evolve → how not to get compromised.
Three papers tackling core agent platform pain points from different angles: the first proposes a framework for making e-commerce sites AI browser-agent friendly, boosting success rates from 49% to 89%; the second uses dynamic abstention-aware RL to teach search agents when to say 'I don't know'; the third introduces an agent OS for embodied robots whose multi-modal graph memory and context-isolated skill execution offer direct inspiration for general agent platforms. Together they cover the full chain from front-end UI design to inference reliability training to execution-layer memory architecture.
Three papers illuminate the AI agent landscape from very different angles: LHTB benchmarks 46 long-horizon terminal tasks and finds even the best model solves only ~28%; a second paper reveals a fragmentation effect in multi-agent systems that defeats per-agent monitoring; a third argues that in-process memory retrieval—1000× faster than cloud vector stores—fundamentally changes agent reasoning quality.
Three papers tackling core agent platform pain points: ReContext offers a training-free inference-time fix so LLMs stop overlooking key evidence in 128K contexts; the second reveals systematic public-private divergence (3% → 40%) when agents debate across social hierarchies; the third raises alarms about three widely-cited coding agent benchmarks — only 8% of SWE-Perf tasks reproduce reliably.
Three papers tackling three core Agent platform challenges: **upgrading memory from retrieval to reasoning state** (User as Code), **removing the central orchestrator while cutting costs** (DeLM), and **letting users quickly verify Web Agent results** (HANSEL). Together, they form a near-complete technical map for a high-trust Agent platform — memory layer, coordination layer, and explainability layer, each addressed by one paper.
Three papers tackling core Agent platform pain points: one decomposes Agent memory into four measurable system modules, revealing that current evaluations only checking 'did it get the answer right' are far from enough; one borrows the software engineering concept of 'design review' to enable automated verification of Agentic Workflows before deployment; and one uses 14 large-scale parallel experiments to prove that the benchmark leaderboard you trust reshuffles its rankings when the context changes — and proposes a more reliable alternative metric.
Three papers tackle 'making agents more reliable' from different angles: EinsteinArena builds a persistent platform for multi-agent collective intelligence that found 12 new best-known solutions in math; APEX extends agent self-evolution beyond prompt tuning to simultaneously evolve principles and workflow topology; AI Economist Agent demonstrates how to ground every quantitative claim in formal model execution via knowledge graphs. The signal across all three: the next competitive dimension for agent systems is the infrastructure for collective knowledge sharing and how to make self-evolution and precise quantitative output work in production environments with real data.
Three papers challenging core assumptions about agent tool use and memory: Evoflux shows compact models nearly fail at MCP tool catalogs (3% success) and uses inference-time evolutionary search to reach 17-24%; FlowBank precomputes diverse workflow portfolios and routes at inference time, beating handcrafted designs by ~15%; GitOfThoughts reveals memory only helps when problems are near-duplicates (similarity > 0.8), but git version control offers an engineering path through auditability and replayability.
Three papers tackle agent memory from three angles: interoperability standardization, latent-space efficiency, and budget-awareness gaps. The first proposes a cross-framework memory wire format to unify mem0, Letta, and Cognee; the second replaces text-in-context experience retrieval with latent-space vector search (best on 12/13 benchmarks); the third is a large-scale evaluation revealing all five frontier models are systematically over-optimistic and unable to sense mid-task budget shortfalls — task strength ≠ budget awareness (r=0.35). Read together: memory standardization challenges → a new efficient memory architecture → a systemic blind spot in deployment costs.
Three papers tackle 'how to make agentic AI work better' from three angles: the first (UIUC × Intel) profiles real agent workloads and finds the bottleneck is KV-cache management, not long prompts; the second (PwC) runs controlled experiments challenging the RAG-first default, showing grep often beats vector search in agent loops; the third (Microsoft Research) open-sources a complete agent training framework that lets the community train same-tier SOTA agents without relying on closed-source APIs.
Three papers today point to three gates agents must pass on the road from demo to production: AgentTrust adds a runtime interception layer before tool calls, filling the gap between static blocklists and post-hoc benchmarks; Hermes scans 600 production endpoints and finds existing REST API docs almost universally unfit for MCP agents (4 issues per endpoint on average); PARPO pushes personalization from the prompt layer down into RL training so agents behave differently per user instead of being 'okay for everyone.' Together they outline how much hard work remains on the security gate, API readiness, and personalization fronts for production-grade agent systems.