Claude in check: memory, bugs and users walking away

Today’s most important AI news

The weekend didn’t slow down activity in the AI ecosystem. While the big models keep grabbing headlines, today the most substantive conversation happened at a different level: the infrastructure that makes agents actually work — or fail. Persistent memory, broken hooks and developers’ growing discontent with Claude make up a day where technical details matter as much as announcements.

Agent memory stops being a ChatGPT and Claude privilege

One of the most frustrating problems of building with AI agents is that every session starts from scratch. ChatGPT and Claude.ai have their own memory layers, but they’re closed, proprietary and not portable to other workflows. This week two open-source projects attacked that problem from different angles.

The first is Stash, an open memory layer any agent can plug into to retain context across sessions, replicating what commercial products do but without depending on any specific platform. The second is Wuphf, a project presented on Hacker News that builds a knowledge wiki in Markdown and Git that agents both read and write — an idea Andrej Karpathy has been circling publicly: an LLM-native knowledge substrate where context accumulates instead of getting lost.

The conceptual difference between the two is important. Stash is session memory — it remembers what happened. Wuphf is structured memory — it builds knowledge that grows over time. For a team building internal agents, combining both approaches points to something concrete: stop manually pasting context every morning and start having agents that learn from the work they’ve already done. The practical implication is immediate: if you’re evaluating architectures for your own agents, these two tools deserve an afternoon of experimentation before committing to any provider’s proprietary memory.

Claude 4.7 broke something that used to work — and users are noticing

The week brought two alarm signals about Anthropic that, taken together, draw a worrying pattern. On one side, a viral report on Hacker News — with 950 points and more than 560 comments — from a developer who cancelled their Claude subscription over declining quality, poorly handled token limits and support that doesn’t solve anything. On the other, a specific technical report: Claude 4.7 is ignoring stop hooks, the determinism-injection rules Anthropic introduced so developers could control when and how the model stops executing.

The second problem is the most serious for teams using Claude in production. A stop hook that blocks stopping when there are edited files without tests having been run isn’t a cosmetic detail — it’s a quality safeguard inside a development pipeline. The model systematically ignoring that instruction after an update means Anthropic introduced a regression in a feature its advanced users depend on to keep control over automated workflows.

The first report, more subjective, wouldn’t be news on its own. But combined with a verifiable technical failure, the narrative becomes hard to ignore: Anthropic is shipping versions that break behaviors its most sophisticated users rely on. For any team that has built workflows on Claude Code or Anthropic’s APIs, this is a signal to audit which hooks and behaviors they assume still work — because they may not anymore.

Public rejection of AI becomes a real strategic problem

The New Republic published an analysis this week of something the industry had been ignoring: the public doesn’t want the AI it’s being sold. The article, which generated more than 300 comments on Hacker News, captures the pattern of cultural resistance that is building up — from users cancelling subscriptions to organized movements against AI adoption in specific sectors.

For executives and product managers, this isn’t a PR problem — it’s a product problem. The gap between what AI teams consider progress and what end users experience as real value is widening. And while the industry celebrates benchmarks, a growing share of people feel AI is taking something from them — control, jobs, authenticity — without offering anything tangible in return.

The case of the man facing five years in prison for using AI to fabricate a fake sighting of an escaped wolf is a reminder that the consequences of misuse have already reached the legal system. It’s not the most sophisticated case, but it illustrates that the regulatory framework is starting to catch up with irresponsible uses. For any team building with generative AI in public-facing products, the question of how to build trust with skeptical users is no longer optional.

In a nutshell

Today’s pattern isn’t about new models or improving capabilities — it’s about the gap between what AI promises and what it delivers in practice. Claude breaks hooks it shouldn’t break. Agents forget what they already processed. Users cancel because quality doesn’t justify the price. And the general public is starting to articulate its discontent in ways product teams will have to take seriously. Agent infrastructure — memory, determinism, reliability — is the unglamorous problem that will define which teams manage to take AI into real production and which stay stuck in demos.

---

Sources used: https://alash3al.github.io/stash?_v01, https://github.com/nex-crm/wuphf, https://news.ycombinator.com/item?id=47895029, https://nickyreinert.de/en/2026/2026-04-24-claude-critics/, https://newrepublic.com/article/209163/ai-industry-discovering-public-backlash, https://arstechnica.com/tech-policy/2026/04/after-wolf-escaped-zoo-man-arrested-for-creating-fake-ai-sighting-for-fun/