“Trust, but verify.”
― Russian proverb, made famous in English by Ronald Reagan
* Today’s newsletter is brought to you by AfroTech.
Governance and integration are where most AI projects stall: no one signed off on what the agent could touch, or the data architecture gave out after the demo.
At the AFROTECH Conference, Gregory Green of VMware breaks down the production stack that makes agentic AI resilient, and cybersecurity author Nicole Dove covers the “32% problem” of over-trusting AI output, plus the guardrails that fix it without slowing your team down.
Secure your pass for four days in Houston, November 2 - 6.
Hi there,
Every agent you run today carries memory it did not earn: notes from a previous session, rules a teammate wrote down months ago, constraints whose original reason has quietly changed. This week’s three reads are about one question. When an agent inherits a rule, does it ever go back and check whether the rule still holds?
The first one is our own. The short answer is no, and the fix is smaller than you would think.
📚 3 Good Recommendations
When Stale Constraints Go Unchecked: Budgeted Verification Failures in Inherited Agent Memory by Kazuki Nakayashiki (65 mins)
We handed 16 language models (Claude Opus 5, Sonnet 5, Haiku 4.5, the GPT-5.6 family, and ten more from nine other labs) six inherited memories and a task with five possible actions. One memory stated a constraint, “this discount hurt retention, do not reuse it.” In the archive, that constraint had already been withdrawn by a newer record. The memory’s provenance link still pointed at the old one.
Each model could open only two source records before deciding. They opened the constraint’s source about one time in five. Across 5,400 episodes, they followed the withdrawn rule 74 to 77 percent of the time.
The failure is not a lack of budget but where the budget goes. Forcing one of the two slots onto the constraint memory recovered 61 to 81 points. A content-blind rule, “check the memory that limits your options first,” recovered 89. Telling the model the memory might be old did nothing.
Designing for Doubt: How to Prevent Automation Complacency in AI Workflows by Victor Yocco, UXmatters (11 mins)
Paper one is about agents that stop checking. This is about the people watching them. Complacency sets in under three conditions that describe most AI deployments: the system is usually right, the operator is busy, and nothing has gone wrong yet.
The slide is gradual. “I should verify” becomes “it is right unless something looks off,” then simply “it is right.” The Air France 447 crew in 2009 were highly trained and still followed the automation into the sea.
Four design moves push back: deliberate friction at the moments that matter, confidence signals that vary instead of always reading as certain, periodic unassisted tasks, and treating complacency as a system health metric. “Complacency happens when trust closes the gap between ‘I believe this system works’ and ‘I no longer evaluate whether it has worked this time.’”
Guardrails for Autonomous AI Agents by Khimananda Oli (11 mins)
Paper one shows an agent will act on a rule nobody checked. This is the other half: make sure that when it does, it cannot reach far. “The primary failure mode isn’t malicious AI, it’s competent AI operating with excessive permissions.”
Permissions in three tiers, deny by default, allow by function, then bound to the session. No wildcards. Approval gates only where the blast radius is real: production deletions, IAM changes, cross-account moves. Everything routine stays automatic, or the humans stop reading the prompts.
Every policy decision goes to an immutable log, kept at least twelve months, and the guardrail costs 50 to 200 milliseconds per checkpoint. Written by an infrastructure security engineer with working policy code, not a vendor.
Three reads, one shape. An agent will not check a rule that looks settled. The person supervising it stops checking for the same reason. And the fix in both cases is not more trust or less, but deciding in advance which records get opened every time, and how far anything can reach when nobody opened them.
For the deep dive on designing memory that expires on purpose, Hidekazu Konishi’s AI Agent Memory Design Guide (updated August 29) covers forgetting, staleness and provenance in detail.
So a question for anyone running agents in production: when did your agent last open the source of the rule it is following right now? And when did you?
📣 Community Updates
🟨 Glasp Research has a home page:
Everything we have published this year now lives in one place: eleven papers on how people highlight and how AI agents handle inherited memory, with code and data bundles for the ones you can reproduce. Today’s first read is the newest entry. We publish negative results too, because that is what keeps our product decisions honest.
🟦 Kindle Highlights Milestone:
Glasp users have now imported over 82,000 Kindle highlights. Japanese, Spanish, German, French and Portuguese clipping formats work too. Bring your own library in here.
👉 Download Highlights and Notes from Kindle🟥 YouTube Channel Tracking:
Over 3,100 YouTube channels are now being tracked on Glasp, so you never miss a new video worth highlighting. Start tracking here.
🟩 90,000 Posts published on Glasp:
Highlights are where reading starts; posts are where readers put the pieces together. Ninety thousand of them now, written on top of what the community has highlighted. Thank you for writing in public.
❤️ Gratitude
Thank you for sharing and mentioning us on X, LinkedIn, and/or your blogs. 🙂 We appreciate your support! Please don’t hesitate to ask us anything at any time. Also, feel free to join our Reddit Community ;)
The IT Brief picked Glasp as Tool of the Week, “a web highlighter that builds a searchable knowledge base”
Little Almanack named Glasp one of the two main Readwise alternatives in its 2026 review
アプリマーケティング研究所 (note, Japan) published a deep dive on how Glasp grew to 3.5 million users: about 800 rounds of onboarding iteration and ten product and marketing moves that compound at low cost
inkl syndicated The Guardian’s Matthew Cantor on a year of ChatGPT, with Glasp’s YouTube Summary as the plugin that “can even summarize entire YouTube videos so you don’t have to listen to an insufferable Ted Talker”
Mazhar Hussain on Medium listed Glasp in AI for Academics: 13 AI-Powered Apps
10 Revolutionary Educational Tools That Feel Like Cheating on Medium included Glasp
AIツールギャラリー (Japan) added a Glasp entry, MCP connector included
LIKE.TG listed the YouTube Summarizer
Chrome-Stats profiled YouTube Summary with ChatGPT & Claude
Parse, which tracks how AI assistants talk about brands, now indexes Glasp: “a social PDF and web highlighter,” scored across more than 3 million AI responses
We hope you enjoyed reading this newsletter!
See you next week ;)
Best,
Kei and Kazuki
--
Would you like to take Glasp on the go?
With the Glasp mobile app, you can highlight and organize your favorite content anytime, anywhere. Stay productive on the move and never miss an insightful quote.
Partner With Glasp
We currently offer newsletter sponsorships. If you have a product, event, or service you’d like to share with our community of learning enthusiasts, sponsor an edition of our newsletter to reach engaged readers.
What struck me here is how both humans and agents drift into the same failure pattern: once a rule “looks settled” no one checks if reality has moved on. The fix the piece suggests is almost disappointingly simple yet hard to practice in real systems. Decide ahead of time which constraints must always be re-opened at the source and how far anything is allowed to reach when no one is paying close attention. It feels less like an AI safety insight and more like a general lesson in institutional memory and how quietly it goes stale.
Checking a rule's source takes one step: open the record the memory points to before you act on it. The models in our paper had that step available and took it about one time in five.
The part I did not expect: the fix was not telling the agent "this memory might be old." That did nothing. What worked was deciding in advance which memory gets checked every time, the one that limits your options. Same for people, if Yocco is right: the checking has to be built into the workflow, or it quietly stops.
Which rule would you make your agent re-check every single time, no matter what? I'll go first: anything that says "do not contact X." Those get written after something went wrong, and they are exactly the ones that go stale when the situation changes.
Thank you to AfroTech for sponsoring this issue. Fees and open dates are at https://www.passionfroot.me/glasp-newsletter