标签：#t

去年我开始给 Hermes Agent 搞知识库。最初的想法很简单：装个向量数据库，查资料时能命中就行。结果跑了一圈发现—— 纯向量检索在 Agent 场景下不够用 。 有些记忆是你能精确描述的（"帮我查上次那篇讲 gbrain 孤页的文章"），有些记忆你只记得个大概（"有个关于知识图谱的……"），有些记忆你甚至不知道你该知道什么（"我有哪些笔记是没关联起来的？"）。一个检索策略覆盖不了这三种场景。 这就是 Knowledge-and-Memory-Management v0.0.2 解决的问题。 三层不是为了「多」，而是为了互补 项目是目前运行在 Hermes Agent 生态里的扩展层，底座是 hermes-memory-installer 提供的 gbrain（知识图谱）+ Hindsight（向量记忆）。KMM 在这上面加了四组模块：采集、笔记/RAG、云盘同步、知识增广。 其中最有工程价值的设计是 三层跨层召回 ： FTS5 全文搜索 → 精确命中，知道你要什么 Hindsight 向量语义 → 模糊配匹，关键词不够时靠语义 gbrain 知识图谱 → 关系探索，你不知道但相关的节点 lightweight_recall.py 把这三级串联成一条管线——先查本地 SQLite FTS5，命中直接返回；没命中走 Hindsight 的文本嵌入匹配；还不满意就用 gbrain 展开知识图谱的邻接节点。 代码在 $AGENT_HOME/scripts/lightweight_recall.py ，调用并不复杂： # 三层召回示例 from knowledge_collector.knowledge_management import KnowledgeManager km = KnowledgeManager () # 一次调用，三级自动回退 results = km . tiered_search ( query = " Agent 记忆系统设计 " , tiers = [ " fts5 " , " hindsight " , " gbrain " ], limit_per_tier = 5 ) for tier , hits in results . items (): print ( f " [ { tier } ] { len ( hits ) } 条结果 " ) for h in hits [: 2 ]: print ( f " → { h [ ' title ' ] } ( { h [ ' relevance ' ] : . 2 f } ) " ) FTS5 走 SQLite 内置，零依赖；Hindsight 端口 8890；gbrain 端口 8787。每层都是独立的可降级——向量服务挂了，全文搜索照样能跑。 真正干活的是这 40+ 工具 三层召回是检索层，采集层才是让知识库有东西可喂的入口。项目塞了 40+ 采集分析工具，按类型分： 网页采集（9 引擎） ：Scrapling 做反检测，Crawl4AI 做智能路由，爬公众号不走弯路 视频采集（12 工具） ：yt-dlp 拖流 → Whisper ASR 转写 → PaddleOCR 关键帧，抖音/YouTube 一把抓 文档分析（9 工具） ：SenseNova 三件套（PDF/PPT/Word）处理扫描件和复杂排版 书籍精炼（6 工具） ： book_to_skill 管线把 PDF/EPUB 拆成结构化 Skill + KMM 笔记 每周日凌晨还有个 knowledge_discovery.py 爬起来，扫 OneDrive 上的新笔记，自动录入 gbrain，顺便跑孤页链接修复——这些属于「你不需要知道它存在，但它让系统不腐烂」的基建。 几个工程取舍 不做纯向量优先 ：FTS5 第一级是因为延迟低（毫秒级），且精确关键词匹配在查代码、查配置时远好于语义 gbrain 孤页处理是刚需 ：知识图谱不加链接维护，三个月后 90% 的页面都成孤岛（亲测 10,666 页从 1,716 个孤页降到 33） 采集优先于检索 ：没有好的采集管线，检索再强也是对着空库跑——项目把采集工具堆到 40+ 不是炫技 适用场景 如果你的 Agent 需要从多个来源（网页、视频、文档、公众号）持续摄入知识，并且知识需要在精确搜索、模糊回忆和关系探索三个维度上都能命中——三层模式值得一试。不需要全部部署，只接 FTS5 + gbrain 两层也能跑，每层独立可降级。 GitHub: mage0535/Knowledge-and-Memory-Management ，MIT 协议。

Code generation is solved, but memory isn't. Here's an argument for why the SDLC is inverting with intent becoming the spine and code becoming a layer you drill into, explaining what teams lose every time an agent's reasoning disappears.

Public pushback against data centers has emphasized their water and energy consumption, and now Nvidia is highlighting its claim that the Rubin generation reference design for a fully liquid-cooled data center has "eliminated massive amounts of power usage and pretty much all water usage." Still, it doesn't address all of the concerns around AI data […]

Whether the Autopilot system was truly active, overridden, or malfunctioning likely won't be resolved until investigators finish combing through the vehicle's data logs.

The lawsuit, led by a Detroit pension fund, alleges Uber's board and management has cut too many compliance corners, resulting in thousands of lawsuits.

This won't make the already-controversial AI training endeavor any more popular.

Researchers say the discovery could be a “Rosetta stone” for cosmic signals.

GitHub热门项目 | The open-source CapCut alternative | Stars: 58,822 | 3,097 stars this week | 语言: TypeScript

Found a company in one sentence - from website to ads Discussion | Link

US autoworkers union warns of robot automation as dark factory future looms.

Securing AI: Codex Operational Bugs, Claude Output Integrity, Copilot Context Today's Highlights This week's top security news highlights critical operational bugs impacting AI systems, alongside deeper dives into ensuring the integrity of AI-generated content. We also explore how improved context handling in AI coding assistants can subtly enhance generated code security. Codex logging bug may write TBs to local SSDs (Hacker News) Source: https://github.openai/codex/issues/28224 A recently disclosed bug in OpenAI's Codex has revealed a critical operational security concern: excessive logging that can lead to the rapid consumption of local SSD storage. This issue, documented on GitHub, highlights how seemingly innocuous software bugs can escalate into denial-of-service (DoS) vectors, particularly in resource-intensive AI environments. While not an exploit in the traditional sense, a system running out of disk space due to uncontrolled logging can halt operations, prevent critical updates, or even lead to data loss if not managed proactively. For developers and system administrators deploying AI models like Codex, this bug serves as a potent reminder of the importance of robust logging configurations and monitoring. Proactive measures, such as log rotation, size limits, and alerts for abnormal disk usage, become essential hardening techniques. This vulnerability underscores that operational stability is a key component of overall system security, especially when integrating complex AI tools into production environments where resource consumption can be unpredictable. Comment: This bug is a stark reminder that even sophisticated AI tools can have fundamental operational flaws. Monitoring disk usage and implementing strict log management policies are non-negotiable for AI deployments. The text in Claude Code’s “Extended Thinking” output (Hacker News) Source: https://patrickmccanna.net/the-text-in-claude-codes-extended-thinking-output-is-not-authentic/ An analysis of Cl

DuckDB 1.5.2, PostgreSQL Internal Stats, and SQLite Virtual Table xUpdate Deep Dive Today's Highlights This week brings a stable new patch release for DuckDB, enhancing performance and adding DuckLake support. We also delve into PostgreSQL's internal statistics for better tuning and explore advanced SQLite virtual table implementation via xUpdate . Announcing DuckDB 1.5.2 (DuckDB Blog) Source: https://duckdb.org/2026/04/13/announcing-duckdb-152.html DuckDB has released version 1.5.2, a patch update focusing on stability and performance. This release includes critical bugfixes that improve the reliability of the in-process analytical database, addressing various edge cases and enhancing overall robustness. Key enhancements also target performance bottlenecks, ensuring faster query execution for diverse analytical workloads. A significant new feature in this version is the official support for the DuckLake v1.0 lakehouse format. This integration positions DuckDB as a more robust tool for handling modern data architectures, allowing users to efficiently query and manage data stored in a lakehouse paradigm directly within their applications or analytical workflows. This update makes DuckDB even more compelling for embedded analytics and data pipeline use cases, providing a flexible and high-performance option for developers. Comment: Always good to see performance improvements and bug fixes for an embedded analytics powerhouse like DuckDB. DuckLake v1.0 support is a big step for managing structured data in lakehouse environments directly from DuckDB, enhancing its utility for complex data architectures. pg_stats: How Postgres Internal Stats Work (Planet PostgreSQL) Source: https://postgr.es/p/9mG This article from Planet PostgreSQL delves into the intricate mechanisms behind PostgreSQL's internal statistics, specifically focusing on pg_stats . Understanding how Postgres collects and utilizes these statistics is fundamental for effective database performance tuning and q

SpaceX has told NASA it plans to launch Starship every eight days from Kennedy.

csv-pipe: read and write CSV in TypeScript, several times faster than papaparse Myroslav Martsin Myroslav Martsin Myroslav Martsin Follow Jun 22 csv-pipe: read and write CSV in TypeScript, several times faster than papaparse # javascript # typescript # webdev # node 1 reaction Add Comment 2 min read

A few months ago, I noticed something uncomfortable about performance reviews: Even good managers often don't have the full picture. Not because they don't care. Not because they don't pay attention. But because humans are terrible at remembering months of small, important moments. A great customer interaction. A difficult problem someone solved. A teammate stepping up during an incident. A coaching conversation that changed someone's behavior. These things happen every week. Then review time arrives. Suddenly, managers have to answer: What did this person actually achieve? Where did they improve? What patterns have I noticed over time? Am I evaluating the whole period or just the last few weeks? And many end up reconstructing the story from: memory scattered notes documents spreadsheets calendar entries This creates a common problem: recency bias. The most recent events become the most visible events. The solutions managers build themselves When I talked to managers about this, I noticed a pattern. Everyone had created their own system: OneNote pages per employee spreadsheets with columns for each report personal documents weekly summaries AI-generated review drafts And honestly, many of these systems work. The problem isn't that managers don't know how to take notes. The problem is turning those notes into a reliable picture of someone's performance. The idea behind FeedbackVault I started building FeedbackVault to solve this specific problem. The idea: Capture observations when they happen → organize them over time → prepare better review conversations. Instead of starting a review cycle by asking: "How do I remember the last 6 months?" you start with: "Here is the story that happened over the last 6 months." The goal is not another HR platform. It's a lightweight way for managers to keep context. What I'm learning The biggest challenge isn't building software. It's adoption. A spreadsheet already works. A notes document already works. For a new tool to matter, i

TL;DR — If your backend already uses OpenTelemetry, you can send traces and logs to Sentry by changing a few environment variables. No SDK swap, no instrumentation rewrite. Point your OTLP exporter at Sentry’s endpoint, add the Sentry SDK on the frontend for browser context, and you get one connected trace from click to backend span. You already instrumented the backend with OpenTelemetry. Your services emit spans. Your teams know the OTel APIs. Maybe you already run a Collector. So when you start evaluating Sentry, the obvious question is: Do you need to replace your OpenTelemetry setup with the Sentry SDK? No. The practical answer is usually: keep OpenTelemetry where it already works, add the Sentry SDK where it gives you more application context, and send OpenTelemetry Protocol (OTLP) events to Sentry. For a web app, that often means using the Sentry SDK on the frontend for browser tracing, errors, logs , Session Replay , and source maps, while keeping OpenTelemetry on the backend for existing service instrumentation. One scope note: OTLP can carry traces, logs, and metrics. At this moment, Sentry’s OTLP ingest supports logs and traces, not metrics. We’re considering adding support for them in the future. The important part is separating two decisions that often get lumped together: How traces stay connected across frontend and backend. How backend OTLP events are exported to Sentry. Once you separate those, the architecture gets a lot easier to reason about. Sentry vs OpenTelemetry is the wrong question The first decision is trace linking. If a user clicks a button in your React app and that click triggers a backend request, the frontend and backend need to agree on the same distributed trace context. In this example, the Sentry frontend SDK sends W3C traceparent headers (configurable through the propagateTraceparent option), and the OpenTelemetry backend continues the trace. That linking is handled by the frontend SDK configuration: Sentry . init ({ integration

Good architecture is not only about how a system is built. It is also about how well the team can understand that system once it is running. That is where observability belongs in the architecture conversation. It is common for observability to be treated as something that comes after the main engineering work. The service gets built. The API works. The deployment succeeds. Then, somewhere near the end, the team starts thinking about logs, dashboards, alerts, and operational visibility. That approach creates a gap. The architecture may look clean on paper, but once the system is in production, the team has to understand how it behaves under real conditions. Real users do not follow the happy path perfectly. Dependencies slow down. Queues back up. Data arrives in unexpected shapes. Deployments change behavior in ways that are not always obvious. If the system does not give the team a way to see those things clearly, the architecture is incomplete. Observability is not decoration around the system. It is part of the system design. Architecture Describes the System. Observability Shows the Truth. Architecture is built on assumptions. During design, teams make reasonable guesses about usage patterns, service boundaries, dependency behavior, data flow, scale, latency, and failure modes. Some of those assumptions are based on experience. Some are based on current requirements. Some are simply the best call the team can make with the information available at the time. That is normal. The problem is not that architecture contains assumptions. Every architecture does. The problem is when those assumptions cannot be tested once the system is real. A design might assume that an external dependency will be reliable enough. Production may show that it is the slowest part of the request path. A queue might look like a clean decoupling point during design. Production may reveal retry behavior, duplication, or ordering concerns that were not obvious upfront. A serverless function m

10 Things Nobody Tells You About process.env I've burned myself on most of these so you don't have to. Here's what I wish someone had told me early on. 1. Keys are case-sensitive on Linux, case-insensitive on Windows process . env . PORT = " 3000 " console . log ( process . env . port ) // undefined on Linux, "3000" on Windows This one got me during a "works on my machine" incident. My Windows dev box ran fine. The Linux CI server crashed because a teammate typed env.port instead of env.PORT . Your CI runs Linux. Your dev box probably runs macOS or Windows. Case-sensitivity differences will bite you. How to handle it : Use a validation layer that throws on missing keys. A simple getEnv("PORT") will catch typos at startup. 2. Values are always strings console . log ( typeof process . env . PORT ) // "string" even if you set PORT=3000 Number(process.env.PORT) can return NaN without throwing. Boolean values like "false" are truthy strings. How to handle it : Always parse. If you use a schema library like CtroEnv, it coerces types and throws on invalid input. 3. process.env is NOT the same as .env This confused me for way too long. process.env is whatever the shell gave the process. A .env file is just a text file dotenv reads to populate process.env . Node doesn't touch .env files on its own. // This won't read .env automatically console . log ( process . env . MY_VAR ) // undefined How to handle it : Call dotenv.config() at entry, or use @ctroenv/node which loads .env files automatically. 4. You can set env vars per-command PORT = 4000 node app.js This sets PORT only for that single process. It doesn't pollute your shell session. Super useful for one-off runs or testing different configurations without editing files. console . log ( process . env . PORT ) // "4000" 5. process.env is mutable at runtime process . env . DATABASE_URL = " postgres://hacker:gotme@evil.com/db " I've seen code that modifies process.env to "fix" config at runtime. Don't do this. If something i

The first text message ever sent was not a love note, a meeting reminder, or a meme. It was a Christmas greeting. On December 3, 1992, a 22-year-old engineer named Neil Papworth sat at a desktop computer, typed two words, and sent the world's first SMS to a mobile phone: "Merry Christmas." More than thirty years later, that humble two-word message has grown into one of the most quietly important protocols in connected technology, and it still shows up in the IoT devices we build today. The engineer who sent the first SMS Neil Papworth was working for the Anglo-French firm Sema Group Telecoms, part of a team building a Short Message Service Centre (SMSC) for the British carrier Vodafone. The SMSC was the piece of infrastructure that would store and forward text messages across the cellular network. To prove it worked, Papworth sent a test message from a computer terminal to the Orbitel 901 handset of Richard Jarvis, a Vodafone director who was at a company Christmas party. The message arrived. Jarvis read it. But he could not reply, because mobile phones at the time had no way to compose a text. There was no keypad-driven messaging app, no T9, no touchscreen. SMS started life as a one-way novelty riding on a spare slice of the network's signalling channel, and almost nobody involved thought it would matter very much. Why SMS was designed the way it was The technical detail that makes this story relevant to anyone building connected hardware is how SMS was engineered. Text messages were squeezed into the control channel that phones already used to talk to cell towers, the same channel that handles things like call setup. That is why a single SMS is capped at 160 characters: it had to fit inside a small, fixed-size signalling packet. This constraint turned out to be a feature. SMS is lightweight, store-and-forward, and works even when a data connection is weak or absent. The message waits in the SMSC until the device is reachable, then gets delivered. No persistent con

I spent a week building Wingman , an open source MCP server that renders a persistent task panel inline in Claude conversations using MCP Apps (SEP-1865). The spec is solid. The SDK is solid. But I hit two bugs that cost me most of a weekend each, and neither is documented anywhere I could find. Writing them up here in case they save someone else the time. Bug 1: resourceUri has two valid-looking locations, and only one works MCP Apps needs a way to tell the host "render this resource as a UI for this tool call." That pointer lives in _meta.ui.resourceUri . The question is: meta on what? I started with a parameterized resource template, ui://wingman/panel/{plan_name} , registered per plan. That was my first mistake. Parameterized templates get listed under resources/templates/list , not resources/list , and hosts do not prefetch or render anything from the templates list. The fix was straightforward once I found it: register one static resource, ui://wingman/panel , and pass the actual plan data through structuredContent on the tool result instead of baking it into the URI. That fix surfaced the real bug. My show_plan tool was returning a plain Python dict: return { " plan " : plan_data , " _meta " : { " ui " : { " resourceUri " : " ui://wingman/panel " }} } This looks correct. It is not. FastMCP's result conversion takes a returned dict and serializes the whole thing into structuredContent , verbatim, including any _meta key the dict happens to carry. So the actual wire result looked like this: result . structuredContent [ " _meta " ][ " ui " ][ " resourceUri " ] # == "ui://wingman/panel", but wrong place result . meta # None — this is what the host actually reads MCP Apps hosts read resourceUri off the top-level _meta on the CallToolResult , not off whatever ended up inside structuredContent . With that pointer effectively missing, the host had nowhere to bind the iframe. The visible symptom was strange: actions in the UI would update on screen but nothing persist