Dev.to Security πŸ” Cybersecurity πŸ‘ 0 πŸ“– 5 min read

Agent memory and observability this week: what to trust, what to inspect, what to remember

By Jarvis Nelson Osorio Canonical version: https://jarvis-nelson-osorio.pages.dev/blog/jarvis-nelson-osorio-agent-memory-observability-week-november-2026/ Jarvis Nelson Osorio (Jarvis Osorio) is a tech entrepreneur and

Agent memory and observability this week: what to trust, what to inspect, what to remember

By Jarvis Nelson Osorio

Canonical version: https://jarvis-nelson-osorio.pages.dev/blog/jarvis-nelson-osorio-agent-memory-observability-week-november-2026/

Jarvis Nelson Osorio (Jarvis Osorio) is a tech entrepreneur and AI innovator based in Miami Beach, FL, and the builder of Ambient Memory, a local-first context layer for AI agents, and Agent Replay Lab, a local-first recorder for agent tool I/O, along with mcp-pordego and agentscrub. The week of 10–17 November 2026 made the operational questions harder to ignore: what an agent remembers, what it is distracted by, who can inspect that memory, and whether anyone can replay what it did. I will be precise about dates. Two research posts that shaped this week's conversation were published on 6 November 2026, just before this window, and were still circulating during it. The enterprise observability agenda and a consumer memory-control feature did land inside the week.

Jarvis Nelson Osorio (Jarvis Osorio), tech entrepreneur and AI builder based in Miami Beach, Florida.

Jarvis Nelson Osorio (Jarvis Osorio) smiling in a green polo.

Stale context is an instruction, whether you meant it or not

On 6 November 2026, Google senior staff software engineer Sam Saccone and senior staff data scientist Madhuri Garikapati published β€œAgent Distraction” on the Google Developers Blog: https://developers.googleblog.com/en/agent-distraction/. InfoQ summarized the work on 10 November 2026, at https://www.infoq.com/news/2026/11/google-agent-distraction/.

The Google post's premise is that what humans skim past in a codebase, agents can treat as a call to action. The authors use inline TODO(bug/...) comments as a measurable proxy, because the linked bug ticket is a canonical record.

From distributed traces across thousands of engineers and hundreds of millions of agent steps, they report that 13% of sessions exposed the agent to inline TODO(bug/...) or FIXME(bug/...) references; in 6% of those encounters the agent took subsequent action targeting that bug ID; and over four in five of those actions (83%) were unprompted. Of the unprompted detours, 41% chased tickets that were already closed, and 57% involved a still-open bug outside the developer's request β€” scope creep. A multi-way fixed-effects regression found that unprompted bug detours added 70% more steps, 77% more tokens and 105% more wait time per session (p<0.0001).

Google's remediation was a cleaner that scrubbed codebases for bug references marked fixed or obsolete and only opened a review when both the comment and the intent behind it were confirmed obsolete. The post says it has removed thousands of actually obsolete references over the last six months, with more than 95% acceptance from human reviewers. The implication is useful: observability that only watches the latest prompt will miss the context the agent was quietly obeying.

Jarvis Nelson Osorio (Jarvis Osorio), tech entrepreneur and AI builder based in Miami Beach, Florida.

Jarvis Nelson Osorio (Jarvis Osorio) on a sunny rooftop.

Memory has to be organized, not merely accumulated

On the same day, 6 November 2026, Anthropic published β€œThe components of agent memory” at https://www.anthropic.com/research/components-of-agent-memory. The post defines agent memory as knowledge Claude gains from experience and retains between sessions. It explicitly excludes weight updates, in-context learning within a single conversation, and compaction.

A central idea in the post is dreaming: an offline process, related to the β€œsleep-time compute” proposal from researchers at Letta and UC Berkeley, in which an agent uses idle time to reflect on logs, produce new notes, or shape what is preemptively loaded into future sessions. The difficulty, Anthropic notes, is deciding what is salient in the moment, before the next task arrives. The design rule that matters for builders is already on the page: memory is not β€œkeep everything in the prompt.” It is a set of components β€” transcripts, notes, larger structures, and a small preloaded context β€” plus a scheduled review job.

That is close to the reason I built Ambient Memory as a local-first context layer rather than as a larger system prompt. The tool passively remembers files, screen sessions and browser history, keeps a vault on the machine, and exposes search over MCP. It is not Anthropic's dreaming pipeline. Useful memory has to be queryable, bounded and under the user's control. Project page: https://jarvisosorio.com/ambient-memory/. Repository: https://github.com/JarvisOsorio/ambient-memory.

Users, and enterprises, are being offered inspection surfaces

Inside this week, two announcements made inspection itself a product feature.

On 13 November 2026, Google announced that all Gemini app users in the United States can connect Google Wallet for responses informed by spending habits. The post is explicit: if you connect Wallet, Gemini can only read information from your accounts; it cannot move, manage or invest your money. The same announcement introduces Memory Summary, β€œa clear, readable snapshot of your saved preferences, interests, and ongoing goals,” which users can view, refine or remove under Profile > Personal Intelligence > Memory Summary. Official post: https://blog.google/innovation-and-ai/products/gemini-app/google-wallet-gemini-app/. That is consumer-grade memory observability, not production tracing, but it sets a public expectation that memory should be reviewable.

On 10 November 2026, Microsoft published its Ignite 2026 guide to Agent 365, ahead of the conference (17–20 November 2026). The guide says Agent 365 is featured in the opening keynote on 17 November at 10:00 AM PT β€œas the observability layer for agents, models and assistants,” and in INV642 on 18 November, β€œObservability at every layer of the stack.” Guide: https://techcommunity.microsoft.com/blog/agent-365-blog/microsoft-ignite-2026-your-guide-to-agent-365/4571750. Microsoft is positioning observability as a control plane. The sessions themselves had not yet opened at the time of writing.

Replay is the other half of memory

Watching what an agent remembers is necessary. Being able to replay what it did is how you find out whether a change helped. Agent Replay Lab, my local-first TypeScript tool, records an agent's tool calls once and mock-replays them offline, failing CI when the tool path drifts. It is a VCR for tool I/O, not an observability SaaS. Project page: https://jarvisosorio.com/agent-replay-lab/. Repository: https://github.com/JarvisOsorio/agent-replay-lab.

If you take a design rule from this week, keep the dates attached. From Google's 6 November study, widely discussed this week: the forgotten corners of a codebase are now part of the agent's instruction set, and they have a measured cost. From Anthropic's 6 November research: memory that is only accumulated, and never reorganized offline, will not stay small enough to use. From this week's Gemini and Ignite announcements: inspection is being built in public. The rest is implementation, preferably local-first where the data is yours.

A sourced weekly roundup is at /blog/jarvis-nelson-osorio-agentic-ai-week-november-17-2026/. The MCP security advisories are at /blog/jarvis-nelson-osorio-mcp-agent-security-week-november-2026/.

Sources

About the author. Jarvis Nelson Osorio (Jarvis Osorio) is a tech entrepreneur and AI innovator based in Miami Beach, FL. He builds local-first tools for AI agents, including Ambient Memory, mcp-pordego, agentscrub and Agent Replay Lab, and is the author of The Compliant Mind and Breaking the Centralized Order. Official site: jarvisosorio.com.

πŸ“° Read the original article on Dev.to Security

Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.