By ThinkIdiot
AI gets longer memory, travel tools and a billion-dollar security startup
Get this by email →
AWS brings memory into agent retrieval. Google adds holiday planning features to Ask Maps. Cybersecurity AI startup Corma reaches a $1 billion valuation in about three months.
News
Build memory-aware AI agents with Agentic Retrieve on Amazon Bedrock Managed Knowledge Base
AWS has connected Amazon Bedrock Managed Knowledge Base with Amazon Bedrock AgentCore memory through a single AgenticRetrieveStream call. The integration brings conversation history and long-term memory into knowledge retrieval. Developers can now use that call to restore a conversation, retrieve memories and save the new turn after generating an answer.
Shared resources, separate access rights
Developers enable the integration by adding memoryConfiguration to their request. The feature is opt-in and backward compatible. Retrieved memories help compose knowledge-base subqueries. The request's userContext limits document results according to the user's access control lists. The example shares one knowledge base and one AgentCore memory resource across employees. It uses the Amazon Cognito sub claim for memory identity and verified email for knowledge-base identity. IAM enforces per-user memory isolation even when application code gets it wrong. (AWS Machine Learning)
thinkidiot take: IAM can enforce private memory even when application code gets isolation wrong. I would make that enforcement the starting point for the shared employee assistant in this walkthrough. Document permissions and memory identity have distinct controls here, so both deserve deliberate configuration. That boundary matters more to me than reducing the workflow to a single call.
New Ask Maps features to help with holiday travel and prep
Google announced new Ask Maps features on November 16, 2026, for holiday travel and preparation. Users can now generate day-by-day road-trip itineraries and search their saved lists. New agentic capabilities also let Ask Maps find local appointments online and call businesses.
The itinerary connects directly to directions
Road-trip schedules include numbered map pins and sharing options. Desktop users can export them to Google Docs. Retrieved itineraries show live traffic, opening status and information about the atmosphere at the next stop. Users can add stops through Ask Maps and tap directly into directions from their itinerary. The “Know before you go” tips cover parking, charging, weather, potential delays and tolls. Appointment searches cover beauty, pet care, auto repair and home repair. (Google)
thinkidiot take: Ask Maps now puts live traffic and opening status inside retrieved itineraries. I would use those details to check stops before starting directions. Adding stops and navigating from the same itinerary gives the schedule a practical role during the trip. For me, that connection is more useful than the ability to generate the schedule itself.
Cybersecurity AI startup Corma reaches $1B valuation in about 3 months
Corma says it has raised a $125 million Series A led by NEA at a $1 billion valuation. The funding comes three months after the AI cybersecurity startup introduced its product. It brings Corma's total funding to $185 million.
Security incidents brought a surge of interest
Corma emerged from stealth in August. Its backers include Sequoia Capital, Khosla Ventures and Coatue. The startup offers AI models and agents trained for cybersecurity defense. TechCrunch reports that interest surged after October 21 security incidents involving PHP, Node.js and LiteLLM. Corma describes itself as a frontier AI research lab focused on protection. (TechCrunch)
thinkidiot take: Corma reached a $1 billion valuation three months after introducing its product. I would evaluate its defense models and agents against the protection work I need done, rather than use that valuation as my selection criterion. The reported surge of interest followed security incidents, which explains the attention without establishing the product's effectiveness. I give the funding less weight than the defensive capability when choosing what to run.
Tech leaders are convinced that Jensen Huang’s exit was a plot against him
The Verge reported on November 16, 2026, that Jensen Huang had resigned from Nvidia after more than 33 years. Prominent tech figures are calling for his reinstatement and questioning the circumstances of his departure. Speculation about a plot against him is circulating, while US agencies had not confirmed allegations that former Megaspeed executive Alice Huang was a Chinese spy.
Cloud export controls enter the dispute
David Sacks questioned the spy allegations. Epic Games CEO Tim Sweeney criticized the circumstances of the resignation, and Replit CEO Amjad Masad called its timing suspicious. SemiAnalysis explicitly labeled its speculation about Huang's removal a conspiracy theory. That theory linked his removal to reduced opposition to the Remote Access Security Act. Huang opposed the act, which would extend AI export controls to cloud servers giving restricted countries access to Nvidia AI chips. The article says the Department of Commerce released an interim final rule on Monday to address that cloud-access loophole. (The Verge)
thinkidiot take: The Department of Commerce released an interim final rule addressing cloud access while the spy allegations remained unconfirmed by US agencies. I would read that rule before drawing operational conclusions from the commentary about Huang's exit. SemiAnalysis itself labeled its proposed connection a conspiracy theory. For decisions about running AI through cloud servers, the rule deserves more weight than speculation about the resignation.
All the questions Nvidia should answer on Tuesday , but probably won’t
Nvidia interim CEO and longtime CFO Colette Kress is scheduled to present third-quarter results on Tuesday. The company has promised an internal investigation into allegations that restricted chips were smuggled into China. Huang was also directed to answer questions in a closed-door session with three Republican-led congressional committees.
Two Megaspeed statements need reconciliation
Nvidia told The Wall Street Journal it blocked Megaspeed from receiving more chips in June 2025. TechCrunch contrasts that statement with Nvidia's October 2025 assurance to CNBC that Megaspeed was legally using the chips for its cloud. Singapore accounted for 22% of Nvidia's revenue in the second quarter of the previous year. After the Megaspeed investigation became public, Nvidia changed its revenue reporting from customers' shipping addresses to their headquarters. A petition calling for Huang's reinstatement had 180,000 signatures. CNBC reported, citing unnamed sources, that Nvidia was not in talks to reinstate him. (TechCrunch)
thinkidiot take: Nvidia says it blocked further chip deliveries to Megaspeed in June 2025, then described Megaspeed's cloud use as legal in October 2025. I would put those statements side by side and ask Kress to explain what each covered. Blocking new deliveries and describing existing use address different questions, so the dates alone do not establish a contradiction. A precise explanation of that distinction would matter more to me than a broad assurance about the investigation.
Trending AI Papers
Ranking source: Hugging Face Papers for 2026-11-17.
VibePrevisBench: Evaluating Agentic Coding for 4D Scene Generation in Blender

An AI coding assistant can write a Blender program that arranges a scene and moves its objects and camera over time. The resulting rough video can guide a video generation model. VibePrevisBench tests whether assistants can turn a brief into scenes that both contain the right details and show them on screen. It also examines where the assistants go wrong.
- Problem: A scene can contain the requested details without making them visible in the finished video. Existing benchmarks do not check both the rendered video and the long scenes with multiple shots that produce it.
- New idea: VibePrevisBench asks an assistant to build a 4D scene, meaning a three-dimensional scene that changes over time, from supplied material. The assistant must provide runnable code and a proxy video, a rough rendering used to guide later video generation. Scene Check runs the code again and checks the scene against hidden reference facts. Screen Check uses a vision-language model, an AI that interprets images and text, to judge whether facts confirmed in the scene are visible in the video.
- Simple example: Think of building a stage set from a director's notes. Putting a prop in the right place is one check; making sure the audience can see it from the chosen camera angle is another.
- Evidence: The benchmark contains 242 tasks across four input types and checks scene facts in nine dimensions. The authors studied 11 frontier coding agents. Understanding space and camera movement separated their performance. Basic staging and camera work were broadly successful, but difficult requests caused failures, and inspecting renders did not lead agents to fix their mistakes.
- Limitation: The abstract gives no numerical scores for individual agents and does not report whether better proxy videos produce better final videos.
- Why it matters: Checking both the scene and its rendered view helps reveal whether an assistant has actually followed a visual brief.
- Paper: VibePrevisBench: Evaluating Agentic Coding for 4D Scene
MIRIX-2.0: Multi-Agent Memory System

AI assistants need access to past information to give consistent answers and reuse useful experience. MIRIX-2.0 divides this memory work among specialized agents while preserving the original records. It aims to find relevant evidence quickly even as the stored history grows. It also turns past task sessions into reusable instructions.
- Problem: Putting memories into a model's trained parameters requires training, while compressing records can permanently discard details. Keeping everything avoids that loss but creates an ever larger collection to search.
- New idea: Six specialized memory agents preserve complete records as the authoritative source. A provenance hypergraph, an index that links several related items together and points back to their source rows, organizes records of events and factual knowledge. Retrieval budgets place explicit limits on how much evidence each question examines. AutoDream, a background process, merges duplicate index anchors without changing the records and turns task histories into reusable skills.
- Simple example: Think of an office that keeps every original document but maintains a separate index for finding relevant passages. Staff can combine duplicate index entries without shredding documents, then turn notes from completed jobs into instructions for future work.
- Evidence: MIRIX-2.0 ranks first among the compared systems on four of five question-answering benchmarks. It scores 91.4% on LoCoMo and 72.6% on LongMemEval-S, leading the strongest baseline on the latter by 14.3 points. With AutoDream disabled, it beats HyperMem by 31.0 points on single-hop document questions and 22.0 points on multi-hop questions, which require combining evidence. It has the highest mean success among compared memory methods on ALFWorld and WebShop, and scores 49.2% on MetaClaw-Bench against 44.0% for MetaClaw's skill library. With AutoDream disabled, it also answers faster than HyperMem on all five timed benchmarks.
- Limitation: The reported speed advantage over HyperMem applies with AutoDream disabled. The abstract does not give the full system's timing or storage costs as its complete record collection grows.
- Why it matters: Preserving source records while limiting each search could help assistants reuse past information without losing details or searching an ever larger amount of evidence.
- Paper: MIRIX-2.0: Multi-Agent Memory System
MMFlash: Faster Speculative Decoding for Vision-Language Models

AI models that work with images and video can generate answers faster by checking text proposed by a smaller helper model. That shortcut loses efficiency when the helper struggles with words that depend on visual details. MMFlash focuses extra training attention on those difficult predictions. It also reduces repeated processing when handling long videos.
- Problem: Fast drafting methods have mainly been developed for text, leaving their use with images and video less explored. The helper's suggestions are rejected more often when predicting the next piece of text depends heavily on visual input, and training on mixed visual and text data does not remove that gap.
- New idea: MMFlash uses speculative decoding, where a small helper proposes tokens, or pieces of text, for a larger model to check together. It measures visual dependence by comparing the model's probabilities for the next token with and without visual input. Training gives extra weight to tokens that rely on visual information and that the helper predicts poorly. For long videos, it uses sparse attention, which limits which parts of the input are considered together, to reduce redundant computation.
- Simple example: Think of a junior caption writer drafting a description for an editor to check. The method gives the junior writer extra practice on words that require looking closely at the picture, where rejected guesses would otherwise slow the work.
- Evidence: Draft acceptance falls from 46% for the quarter of tokens least dependent on visual input to 22% for the quarter most dependent on it. That gap persists after multimodal training. Across multimodal benchmarks, MMFlash reports decoding speedups of up to 3.58 times without changing the target model's output distribution. Reported speed gains over EAGLE-3 and DFlash reach 195.4% and 17.1%, respectively.
- Limitation: The speed figures are maxima. The abstract does not give average gains or identify the hardware and input conditions behind those peaks.
- Why it matters: Better guesses about visual details let image and video models generate answers faster while preserving the target model's output distribution.
- Paper: MMFlash: Faster Speculative Decoding for Vision-Language
Trending AI Repositories
Ranking source: GitHub Trending.
NVIDIA/SkillEvaluator
SkillEvaluator is a Python project for assessing AI agent skills and plugins. It looks at security and compatibility as well as quality, giving readers more than a check of whether a skill works.
- What it is: It checks skills, detects overlapping capabilities, generates evaluation data and measures how skills affect agent behavior.
- What it does: Multi-tier framework for evaluating AI agent skills and plugins with quality gates, semantic overlap detection, synthetic evaluation dataset generation, and live agent evaluation that measures how skills affect agent behavior.
- Who it helps: It helps people building agent skills and plugins. They can assess compatibility across agents and examine how their additions affect agent behavior.
- Limitation: The README lists Python 3.12, 3.13 and 3.14.
- Repository: NVIDIA/SkillEvaluator
google-research/paper-orchestra
PaperOrchestra is a Python project that takes ideas and experimental logs into the paper-writing process. Its focus is on working from unconstrained materials before a draft exists.
- What it is: PaperOrchestra turns ideas and experimental logs into LaTeX manuscripts using agents for outlining, literature review, writing, refinement and plotting.
- What it does: PaperOrchestra is a multi-agent framework for automated AI research paper writing.
- Who it helps: It helps researchers with ideas and experimental logs to write from. They can use those materials as inputs to its writing workflow.
- Limitation: The supplied README excerpt does not show how the resulting papers are checked.
- Repository: google-research/paper-orchestra
kajisho5/ffmpeg-skill
ffmpeg-skill gives coding agents a way to work on footage and check the result. That output check is a useful part of its editing workflow.
- What it is: It provides 42 local FFmpeg tools for coding agents and MCP clients. The tools cut footage, add captions, reframe video, adjust loudness and export for different platforms. They require no cloud service or API keys.
- What it does: Local video editor for coding agents: 42 FFmpeg tools with a machine-readable contract and an MCP server, for Claude Code, Cursor and Codex. Cut, captions, reframe, loudness, platform exports. No cloud, no API keys.
- Who it helps: It helps people who want their coding agent to edit footage. They can have the agent perform edits and then check its own output.
- Limitation: It relies on local FFmpeg to edit footage.
- Repository: kajisho5/ffmpeg-skill
Sources
- 01Hugging Face Papers · Hugging Face Papers
- 02GitHub Trending · GitHub Trending
- 03Build memory-aware AI agents with Agentic Retrieve on Amazon Bedrock Managed Knowledge Base · AWS Machine Learning
- 04New Ask Maps features to help with holiday travel and prep · Google
- 05Cybersecurity AI startup Corma reaches $1B valuation in about 3 months · TechCrunch
- 06Tech leaders are convinced that Jensen Huang’s exit was a plot against him · The Verge
- 07All the questions Nvidia should answer on Tuesday , but probably won’t · TechCrunch
- 08VibePrevisBench: Evaluating Agentic Coding for 4D Scene Generation in Blender · arXiv
- 09MIRIX-2.0: Multi-Agent Memory System · arXiv
- 10MMFlash: Faster Speculative Decoding for Vision-Language Models · arXiv