Briefing
Episode 18 · Primer and discussion · About 14 minutes in two parts
Download the MP3 · Linda and Leonard, AI voices, read a script written from the sources below.
Download the MP3 · Linda and Leonard, AI voices, read a script written from the sources below.
Written from the project’s own code and documentation, the creator’s video and OWASP, and checked on 23 October 2026. Nothing was installed or run. How the ChatGPT connection works, any independent security assessment of the hosted service and default retention periods were not documented. Check the original before quoting.
Before we record, every claim the hosts make is checked against the sources (this table was last updated 23 October 2026). Confirmed means a primary source or two independent reports; Reported means one party says so; Partly right, Disputed, Unverified and Not yet known mean what they say, and the hosts say so on air. This is the table as we left it.
Sign up for our podcast at agile-insights.com.au.
Linda Here is the idea for today. When your AI agents share one memory, every agent's record of your work becomes every other agent's input. So memory is a data flow to govern, not a convenience setting. By the end you will know what agent memory is and the four places it can live. You will know what a popular open source tool called Claude Mem showed this week. And you will know the questions to ask before anyone switches it on.
Leonard Let me set the scene, because plenty of people have never watched a coding agent work. If you have used one, such as Claude Code, you know it forgets. Each session starts fresh. Whatever it worked out yesterday about your project is gone, unless someone wrote it down. On the twenty first of October, a show called The Next New Thing ran a short video. Its guest was Alex Newman, who built Claude Mem. He showed work done in Claude Code being picked up in ChatGPT, with Grok Bot reading the same memory. Newman calls his aim a universal memory adapter for all agents.
Linda That is quite a goal.
Linda Before we get to sharing, let us build up what a memory tool does, in four steps. Step one is capture. The tool watches the agent's session and records what it did: the files it opened, the commands it ran, what you asked for. Step two is compression. A raw session is far too long to replay, so the tool sends it to a language model to boil it down. Claude Mem calls that model the observer. The result is short notes, such as fixed the login bug by changing the timeout. Those notes are the memory.
Leonard That second step is the one I would not have guessed. A model is reading my whole session in order to write the summary. When I picture memory, I picture a file sitting on my laptop. But if a model has to read the session first, I need to know where that model runs. Is it on my machine, or is it somewhere else?
Linda It depends on the install, and that is the point of today. Take steps three and four first. Step three is storage: by default, Claude Mem keeps the notes in a small database on your own computer. Step four is recall: when a new session starts, the tool searches the notes and feeds the relevant ones back in. Now, the model. For Grok Bot, the project's own documentation says the installer picks its hosted service as the observer by default. Running the model locally is something you opt into. The main installer asks you to sign in, which starts a thirty day trial of that hosted observer.
Leonard So the four steps each have their own location. Capture happens on the laptop. Compression might happen on the laptop or in the vendor's cloud. Storage is local by default, but there is a cloud sync option. Its documentation warns that sync uploads your notes and the full text of your prompts...
Linda Which is a lot.
Leonard It is. Recall can come from local storage or from a hosted server, which the documentation itself calls a beta. That is a lot of moving parts behind a one line install command.
Linda Now the sharing. Claude Mem has installers for agents from many different companies: Claude Code, Codex, Gemini, Copilot, Cursor, Grok Bot and more. When several of them point at one memory, a note written by one agent becomes context for another. In the demo, the host had Claude Code review a batch of pull requests. Then he asked ChatGPT to turn that work into infographics, and ChatGPT already knew what Claude had done. Newman's pitch is that you use the best model for each job, with no handoff work.
Leonard Tempting, I must say.
Leonard I suspect the trap for a newcomer is a single word. The video says you can keep memory local or run your own server. Someone in a meeting hears local and writes down private, and the conversation moves on. Is that the misunderstanding you would warn people about?
Linda It is the most common one. Local here usually describes where the notes are stored, not where they are processed. The compression step sends session content to a model, and that model may be hosted. Sync sends prompts to the cloud. And with shared memory, the next agent to read your notes may belong to a different vendor. So the statement in the video is true for storage. It is only part of the picture, and the other parts are where the data actually travels.
Leonard Only part of the picture.
Leonard Then bring it back to the person listening, who governs AI in an organisation rather than writing code. What does all of this mean for them?
Linda Three things. First, a memory tool is a data processor and a supplier. In a client's codebase, the notes include client material, so where each step runs is a data handling question...
Leonard Client material. Right.
Linda Second, defaults are decisions. A hosted model chosen by an installer is a supplier choice that nobody signed off. Third, memory persists, so it is also an attack surface. The Open Worldwide Application Security Project, OWASP, lists memory and context poisoning among its top ten risks for agentic applications. A bad note written today can steer an agent next week.
Leonard Next week. Not today.
Linda So here is what to do. Put memory tools on your register of AI tools and suppliers, each with an owner. For every install, answer four questions. What is captured, who compresses it and where, where it is kept and for how long, and which agents and people can read it. Treat the installer's defaults as decisions to make on purpose. Scope memory by project or client, give it an expiry, and know how to delete it. And decide which vendors may read a shared memory before anyone connects a second agent.
Leonard That is the groundwork. In part two, we take the case for shared memory seriously, and test it against what is and is not known.
Linda The idea, again. When your AI agents share one memory, every agent's record of your work becomes every other agent's input. That makes memory a data flow to govern, not a convenience setting. In this part we test that idea against the strongest case for shared memory. We look at what the evidence settles and what it leaves open. Then we finish with what an organisation should do.
Leonard A thirty second recap for anyone starting here. A memory tool like Claude Mem works in four steps. It captures what an agent does in a session. It sends that to a model, which it calls the observer, to compress into short notes. It stores the notes, locally by default. And it recalls the relevant ones into the next session. Each step can run on your laptop or in a vendor's cloud. Shared memory means agents from different companies read and write the same notes.
Linda The question for this part is simple to ask. Is shared agent memory a productivity win that needs a little hygiene? Or is it a new data path that needs a decision before anyone installs it? Leonard, you take the case for it first, and make it properly.
Leonard Happily, because the case is real. Anyone who uses two or three AI tools knows the handoff tax. You explain the project to one tool, then explain it again to the next.
Linda We have all paid that one.
Leonard We have. In the demo, the host finished a review of pull requests in Claude Code. Then he asked ChatGPT for a visual summary, and it already knew the work. Newman's design choice matters too. He built the observer to run in the background, so nobody has to learn a new way of working. He says a tool you have to learn is a gamble on your evening. People will adopt tools like this precisely because they ask nothing of them.
Linda I agree with every word of that, and it is exactly why it needs a decision. A tool that asks nothing of people also asks nothing of governance. Look at what crosses a boundary. Work recorded by Claude Code becomes readable by ChatGPT or Grok Bot. In a client codebase, that record includes client material. So a second vendor now holds client context, and nobody chose that supplier...
Leonard Nobody chose it. Hmm.
Linda The video shows the result, but not how the ChatGPT side was connected. The project's documentation has no ChatGPT guide that we could find. So nobody can yet say whether that path runs through the hosted service.
Leonard Fair, but let me push back once. Is this really new? Organisations already send client data to cloud services every day. They have supplier reviews, contracts and data processing terms for exactly this. Surely a memory tool is just one more service to put through the same process. If the process works for email and file sharing, why would it fail here?
Linda The process is right, and it does apply. What differs is how the supplier gets chosen. For email, someone signs a contract. Here, a developer runs one install command, and the installer picks the supplier. For Grok Bot, the documentation says the hosted observer is the default, and a local model is opt in. The main installer asks for a sign in that starts a hosted trial. So the same review process applies, but it has to catch a decision that was made by a default, often on one laptop.
Leonard So we agree on the remedy. Treat it like any processor, and make the default a deliberate choice. Then here is my second objection. The tool can be self hosted. Newman says you can run your own server, even in Docker, and keep it isolated. Doesn't self hosting close most of this?
Linda It closes storage, and that matters. It does not decide which vendors' agents read the memory, or which model does the compression. Look at where the controls sit. In the open source version, the main privacy control is a private tag that the user has to type around sensitive text...
Leonard Every single time.
Linda Every time. Automatic redaction of secrets exists, but it is off until someone turns it on. The project's own licence notes reserve several features for the commercial tier. They include team sync, single sign on, role based access, an audit interface and data loss prevention.
Leonard That is persuasive. A control that depends on someone remembering a tag is a habit, not a control. So self hosting is necessary for some organisations, but it is not sufficient. You still need the decision about who reads it, and you still need the controls. Let me add the security angle, because shared memory makes it sharper. The security body OWASP lists memory and context poisoning among its top ten risks for agentic applications. A false or malicious note planted in memory quietly steers later behaviour. With shared memory, one agent can write the note and a different company's agent can act on it.
Linda Which brings us to the short passage on what is confirmed. Here is what the project's own code and documentation confirm. We know what the tool captures, and that storage is local by default. We know the Grok Bot observer is hosted by default, and that sync uploads full prompts. The hosted server is described by its own docs as a beta. Not yet established: how the ChatGPT connection works, and any independent security assessment of the hosted service. Newman says the paid cloud is secure. We found nothing that independently tests that. Default retention periods are also not documented.
Leonard That is a real gap.
Leonard Where does that leave the evidence? It settles the shape of the risk. We know memory is captured automatically, compressed by a model that may be hosted, and readable by agents from several vendors. It does not settle how safe any particular deployment is. That depends on the install choices, which the organisation can control, and on a hosted service nobody has yet assessed in public. The sensible stance is neither to ban it nor to wave it through.
Linda Agreed, on both counts.
Linda So here is what to do. First, put memory tools on your register of AI tools and suppliers, each with a named owner. Second, for every install, answer four questions. What is captured, who compresses it and where, where it is kept and for how long, and which agents and people can read it. Third, treat installer defaults as decisions, and pin versions so an update counts as a change. Fourth, scope memory by project or client, give it an expiry, and know how to delete it. Fifth, decide which vendors may read a shared memory before anyone connects a second agent.
Leonard And if you only do one thing this week, ask your developers which memory tools they already run. The answer is the start of your register.
Linda If you'd like to know more, sign up for our podcast at agile-insights.com.au.
Leonard Thanks for listening.
All of the briefings (38) · Follow Briefing in Apple Podcasts or any podcast app
Everything from Agile Insights: briefings, Making You Think and the reference series