# Portable Agent Memory: Solving AI Vendor Lock-In

**Podcast:** Chain of Thought | AI Agents, Infrastructure & Engineering
**Published:** 2026-10-07

## Transcript

Almost 85% of memories that get written are accessed and read back over 30 days later.
But if I decide I want to switch models for any type of reason or try a different type of harness, I shouldn't have to start over.
No one has to choose.
And where they start doesn't need to be where they end.
A controlled study published three weeks ago moved agent memory between models and found something uncomfortable.
The notes your agent wrote about you can swing its accuracy by 13 points, or sometimes more, in either direction, purely based on which model reads them back.
We need memory to be portable to avoid locking.
Welcome to Chain of Thought.
It's great to see you all.
I am your host, Connor Bronsden.
Joining me today is Kimberly Logan.
Kimberly is head of product at the Walrus Foundation.
Huge shout out to Walrus for sponsoring this season of Chain of Thought.
It's been a pleasure having them.
And Kimberly is building the memory layer on top of a decentralized storage network that's been running in production since March of 2025.
Kimberly, so good to have you on.
Thanks so much for having me, Connor.
I'm really excited to talk about what we're building and why we're building it.
Same here, because I think...
agent memory is such an important part of the ecosystem today.
But let's start by setting the stage a bit.
Kimberly, why is agent memory so important?
And what prompted Walrus to take it on as a memory product?
So agentic memory is important because context is what makes data valuable.
And while we have seen the companies that are creating the models themselves, building longer and longer context windows, that is only as good as when you are in that provider for the context window that you have.
Those million tokens are going to get you pretty far.
But if I decide I want to switch models for any type of reason or try a different type of harness, I shouldn't have to start over.
And as you rightfully pointed out at the top of this, the context swings how valuable your data becomes when you switch.
And so we sat down when we were thinking about how do we address this problem?
Because we heard it from our builders, from our users, and from our partners, that this is where they were getting stuck.
People wanted to be able to use agents and use the technology, but they were overly constrained.
When we looked at the unique properties of the protocol today, we were sitting on essentially the perfect gold mine for solving this problem.
Because it is built on top of a decentralized storage protocol, we were able to make that infrastructure useful to people by building Walrus memory to allow them to carry their agents' memories with them.
and remove the concept of them being locked into whatever provider they started with.
Quick disclosure, Walrus Memory sponsored not just this season, but obviously this episode.
I think this is a very apt moment to be talking about this because, you know, as we're recording this, we're a couple days out from some big announcements for Walrus that we're going to be talking about a little later in this episode.
And we've also...
seen multiple major model announcements pop up where Cloud Opus 5.5 and then Sonnet 5.5 both just dropped in the last week and a half.
And today, as we're talking, OpenAI Dev Day is happening.
It sounds like Sol 6.1 is accidentally dropping while they have also delayed Astra 6.1.
So depending on those models, you may want to adjust where you're spending time.
You may be really excited by DeepSeq v1 Flash and want to be using a more open layer.
And if you have gotten locked into the ChatGPT system or the Anthropic ecosystem or another ecosystem, there is just simply data that you're going to lose and your models will not be as accurate when you try to switch over.
Like, yes, I know there's some import features that have been set up, but there's so much rich context and memory that is set.
kind of locked into these systems.
Yeah, I completely agree with that.
And the problem is not solved with an import function either, because that is a point in time solution.
That is not how real workflows happen.
And that's not how we see our users and builders trying to create today.
They want to be able to use different models for different use cases and different purposes.
They're not setting something up and then moving it over and then continuing on.
I personally use different models for different types of work just in my own day-to-day.
And we also know that our users want to be able to take the Walrus Memory solution and embed it into their own products with their own agents that they're creating so that no one has to choose.
And where they start doesn't need to be where they end.
Totally.
We've talked to folks on this show at the enterprise level who, you know, for example, Fergal Reid at Intercom and FinAI talked to us earlier this year about basically getting off of OpenAI for certain workflows and saving $250,000 a month by switching it to open models for these different workflows that he was running for their customer support agents.
But when you do that, it's a ton of work.
And as an individual dev, I see this as well.
I'm using Codex, I'm using Cloud Code, but I've also got a cheap cursor subscription right now.
I've got an OpenCodeGo subscription.
I'm using OpenRouter for some stuff.
I've got Jev running as a classifier on some of my emails.
There are all these models and classifiers that are interacting.
I need to be able to easily port information between them.
And this gets down to the value prop that I think Walrus Memory is providing.
There is quite a bit of conversation around the primitives here.
Why aren't you positioning against memory companies like NemZero or LangMem?
It feels like Wallace Memory is more about the storage primitive here.
While we did build on top of the storage protocol, I do not believe that we are primarily solving a pure memory problem.
And the opportunity space for us is much bigger than memory alone.
And so I don't consider it a apples to apples comparison.
So I don't publicly benchmark or position it in that way because we are building and see opportunities to create more ways for people to own and use their data.
and share it with others in the way that works for them and to allow companies to verify what that memory and that data is across company boundaries.
And that goes way beyond simply managing your memory or creating better harness plugins to allow you to move beyond coding, different coding agents.
And in fact, a lot of our users aren't even using it to build coding agents.
It's only maybe 30% of the actual usage that we have is for switching coding agents.
People are using it for everything from creating long running study aids to building trading agents that trade on their behalf and build portfolios to tracking.
security incidents and comparing production snapshots.
And so I think that our opportunity landscape is much broader than some of the more targeted memory offerings.
I would love to understand that as a bit more.
Where do you see this infrastructure layer that you've built with Walrus Memory going in the coming months?
So what do people need?
to feel confident about where their data is and where it's going.
First of all, they need to know that it's going to be owned by them forever.
They need to control it.
They need to choose who accesses it and where, and they need a way to verify it.
And it needs to work across trust list boundaries because as we look to a more agentic world where agents are going to be doing more and more things on our behalf, The things they can do today, we would like them to be able to do more, but we need to be able to do it in a way that we feel confident that it's safe and we can prove what happened as well.
We're building the infrastructure to make that possible across many different industries and use cases.
And so concretely, we're going to be thinking about building more entry points for people.
continuing to make it more performant and more robust and adding in additional verification functionality that's workflow based because we're seeing people deploy agents in a workflow context for enterprise versus just more one-off types of use cases.
The vast majority of our users are using our products for long running workflows as well.
Almost 85% of memories that get written are accessed and read back over 30 days later.
And so we feel confident that we're building for long-running consequential workflows, not just spin-up, tear-down types of use cases.
This is obviously a through line for the show, this idea of having great data, curating that data.
and ensuring that you build this contextual moat.
You know, Richmond Alake came on the show this year and called agent memory the last battleground in the AI stack.
Jerry Liu from AlmaIndex argued that the frameworks era has ended because context is the new moat.
You know, we've been talking about context harnesses recently.
And I think it's fascinating to see how Walrus is enabling this evolving contextual memory harness.
for its users.
Can you maybe talk me through a bit of an example of what one of those longer running workflows looks like when it is interacting with the primitive of the memory and data storage that you developed?
Sure.
So I'm going to select some of our agentic trading workflows that we're seeing.
So people want their accumulated knowledge to keep working for them because rebuilding things costs time, energy, and attention.
And so what we are seeing is people embedding walrus memory in with their agents from the day that they creating them, sharing, using that delegate key to accumulate knowledge for that agent.
And then as the agent goes on, throughout its activities of trading or gathering other information or the other business rules that it has been asked to do, it logs that data encrypted with those memories back to Walrus through Walrus memory.
And so we're seeing people's agents become more intelligent over time and remember things more efficiently.
And so in cases...
where we even have a customer that's offered this embedded knowledge to get agents started out of the box, we're seeing it become cheaper for agents to operate and to do so in a more accurate way because they are learning from other agents as well as business knowledge as it's long running.
And so actually it becomes more and more valuable the longer that it runs.
In a case where we have something like security instances, we have a user that has embedded in their day-to-day security workflows.
And so they are going to be taking snapshots of what's happening in production, comparing it with code changes, and using it to verify what's changed to help them do their security analysis.
coding more efficient.
But any time that you need something to accumulate knowledge over time and understand how that knowledge has changed and what's applicable and what's not, we're going to get more value from it.
Season four of Chain of Thought is also presented by Walrus Memory.
I spent much of this year talking to guests about agent memory, and we're all having challenges that go beyond simple session handoff.
You explain your code base to one AI coding tool.
You get it tuned to what you need.
But when you look to hand off to a new model family, you find yourself having to confront the same memory and context problems.
Walrus Memory is solving agent amnesia with a portable memory layer for agents.
Walrus Memory takes your agent's memory anywhere you need it.
The context your agent builds can be read back later from a different app in a different runtime, weeks out, and you set the rules for who reads and writes it.
Python and TypeScript SDKs with native MCP support.
I'm glad to have them behind season four of Chain of Thought because this is a key problem confronting AI builders in 2026.
Learn more at walrus.xyz slash COT.
Kimberly, how is the Walrus team thinking about the verifiability and trust side of this extremely important data?
So verifiability is key to being able to answer the question of what happened and when.
And for us to be able to trust agents with our data and these AI systems with our data, that's a critical question we have to answer.
Right now, AI feels a lot like a black box.
We even see that with a lot of the news stories that have come out.
The open AI agent does XYZ thing messing with the government.
We have the hugging face incident that we remember from earlier in the summer.
And so if you're putting your data in, you need to be able to be able to answer the question of what happened to it.
And so that's why the verifiability piece of it is so important to us in our value prop, but also in terms of making this product available to people because It's hard to get those questions answered, but if you are using Walrus memory, you can because the attestation lives on the blockchain and because you control the delegate key and access to it, you can choose to share that with anyone else that you want to.
Yeah, I think this verifiability piece is really important because memories may be true at a point in time.
But as the context of the business changes, as my personal context changes, as the world changes around us, we may need to adjust the decisions we've made around them.
And data is constantly evolving and growing.
And just because a fact was true back in March doesn't mean that, let's say, GPU prices are the same as back then.
So this verifiability piece is crucial, I think, to allow scalability for this kind of underlying data primitive.
I completely agree with this.
And if I even think about what has changed in terms of recall ability in, in six months of my own life, like my preferences now, and even where I see the product going now are different now after I've, I've had some months of, of it being in market.
And so that's Walrus memory and what we're building doesn't just write the memory.
to the blockchain.
It does that, but we also have a knowledge graph layer that provides the logic to make sure that the recall is against the freshest memories.
All the memories are there, but we also think that it needs to be usable too.
And so while we don't publish memory quality benchmarks, to me that is a product choice because I believe that good recall quality, fast latency, all of that, that isn't something I should be out there bragging about.
It's table stakes that anybody should expect to make a quality memory product work for them.
And so I prefer to let the product speak for itself in terms of the usefulness of it.
And I believe that to be competitive in the market, it needs to be useful for people, not just portable, But people need to be able to win through expertise, workflow, and results, not through necessarily benchmarking competitions.
And I have found a lot more value in keeping the product team focused on what people are trying to do and how to make it more usable for them versus necessarily focusing on what the latest benchmarking quality measuring contest looks like.
Yeah, the models are going to keep hill climbing across the variety of benchmarks.
We need to have the domain expertise and the contextual harness to actually ensure this is successful when you're trying to apply it to enterprise workflows and elsewhere, to your point.
And something else you said that I want to just drill down on very briefly, and this is a bit of a sidebar, but I love that we're seeing blockchain as a technology.
come to fruition, this digital distribution of data and verifiability, I think has been so valuable for AI in particular because the scale of how much data we're creating and the autonomous nature of agents means that this, you know, basically supply chain verifiability that can now be applied to the knowledge supply chain.
And there is so much data being created on the internet today.
And that's just...
just going to continue to 10x as we have more and more agents out there doing things for us that we need to track and understand and verify.
And so to me, it's very exciting to see the promise of this technology that has been talked about 10, 20 years ago now come to fruition.
And it makes me wonder how Walrus is thinking about the fusion of all this digitally created agent data and then the human data that you are also bringing in.
We talked a bit earlier about the access leveling, but how are you thinking about this data infrastructure in the kind of next era?
So I will start by saying that this is the first blockchain company I've ever worked at.
I come from traditional software and I never thought I would work at a blockchain company.
But the reason that I came to work at a blockchain company and build data infrastructure was because after spending almost 20 years working in traditional software, I saw concerns with how things were going, especially around things like what happens to our data after we put it out there, how much AI-generated content is taking over the internet.
we're forging ahead with all of these really exciting ambitions and plans and technology to power an agentic future, but where's the other side of this?
And in the past, we might have been more comfortable putting our data into a centralized system in order to gain connection with people.
We put all of our data into Facebook so that we could connect with our friends.
We published our whole professional history on LinkedIn so that I could meet other like-minded people.
But I wanted to solve the problems that I was seeing becoming more and more prevalent.
And so I came to Walrus to solve those problems because I believe that decentralized technology and infrastructure is what is needed to make AI actually valuable in a human-driven world and to make sure that human people have control and autonomy in it, which I think is so critical, especially because things feel so out of control in a lot of ways.
Like, we don't know what's happening.
We don't know what's being deployed or where it went.
So I wanted people to have a sense of control, and people also includes businesses, control over what makes their business run, what makes their proprietary data valuable to them, so that people didn't have to choose between taking advantage of new technology and knowing what was being done with it.
And Kimberly is not going to brag about it herself, but I will for her when she says, oh, you know, it's the first blockchain company I've worked at.
What she means is she spent seven years at Google before this, alongside other pieces of her career as well, but let's just focus on the Google piece, where she led Google Search's client infrastructure program management organization and had to understand how reams of data would come together.
So I think it's awesome to see folks like yourself who have such a deep background in scaled data saying, okay, How do we solve this problem for the future?
We know that agentic data is only going to continue to increase, assuming that we continue down this pathway, which, given the value of AI in so many areas, it seems very clear we will.
And as the simplicity of using something and the price of it goes down, the actual usage massively increases.
And we see this with roadways, right, where.
You know, when we have three lanes on a highway, it's full, but then we have six lanes.
We're thinking, oh, great.
We're now all that space.
No more people drive because now there's more space.
And we see this, I think, with dev and classifier models suddenly become much more useful.
You know, if great, it's it's so much cheaper right now, but that means we're going to use classifiers on everything.
And that just means more and more data being created, more stuff that we want to track, more need for verifiability, more need to be able to control our data.
And it seems pretty clear to me that a tamper evident memory solution, like what Walrus is building on the data side, can be crucial for establishing databases that have good access logs in the long term, especially as agents start to run amok a bit.
And we can look at all the examples this year to point to why this matters.
Yeah.
And we believe that what we're doing goes even beyond what a good access log can provide.
Because there's many solutions already out there available for access logs.
But we believe people should have a record that other people can check, that other companies, regulators, users, and to be able to compare contents against an earlier commitment to detect changes.
We also think though that unchanged doesn't necessarily mean true.
So integrity doesn't establish accuracy, authorization, or complete capture.
So we do think though that we want to be able to have the evidence in order to have, frankly, an open and transparent way to even be able to look at what happened and when.
And I think an exciting part of this is what you're going to be announcing today when this episode comes out, which is the new Walrus Console.
Can you tell us a bit about what Walrus Console is before I spoil the surprise?
I'm so excited for Walrus Console.
It's solving problems that I've been hearing from builders and users since day one.
Walrus Console is all about...
bringing these awesome properties that we just have been spending the last minutes together talking about to more people.
So the properties of portable, verifiable, programmable access for your data to more people through a simple to use interface or MCP connection to be able to manage your agent memories, your files.
and your other information all in one place.
And Walrus console is not only about what you can do today, but it will be the front page into Walrus for all of this other functionality we're going to continue to build in the future.
And all the workflows we're going to begin to enable all of the different ways that you're going to be able.
to have your agents interact with other agents.
And so we're excited because this is the first step in a long journey of making this technology accessible to more people because I think for a long time, a lot of great blockchain technology has not been accessible to most people.
It was just too hard.
It's hard to figure out how to configure it.
how the blockchain technology itself and how it operates blobs and their exploration and all that, that that's a very high bar to clear and requires lots of education to people.
That is frankly a lot to ask.
And we don't think it should be that way anymore because we think that this infrastructure should be available to everyone.
And you shouldn't have to have complex knowledge regarding, how the underlying infrastructure is configured or works, it should just work for you from day one.
And if I understand it correctly, you know, console lets any developer who's using cloud code, codecs, et cetera, to build with the walrus data primitive through an API, through an MCP, inheriting the kind of three default properties of portability.
privacy and verifiability that we've been talking about throughout this conversation.
And it's essentially designed to supercharge the ecosystem of builders around Walrus.
Yes, that's exactly right.
And we are going to be sharing more news about exciting partnerships and new use cases that are building on Walrus that are powered by this new technology in the coming weeks as well.
What other key takeaways should folks have as they think about potentially building with Walrus Memory and leveraging it in their tools?
I would encourage people to use Walrus Memory beyond their coding agents.
The coding agents, an obvious use case, it's built for that.
We have treated that as a first class.
opportunity, but where we have seen people get unique value from the wow moments that we've heard both from actual end users as well as partners that have reached out to us is realizing that it can be deployed anytime that your data not forgetting what happened and accumulating over time is beneficial.
which means it can solve a super large variety of use cases.
And I would encourage people not to be limited to it and what seems immediately in front of them.
And the second is that it's a product of a very active development.
And so the feedback from both our partners on the small and medium business as well as enterprise side, as well as our very active user community.
Their feedback directly informs what we build next.
And so I would encourage people to participate in our community activation events, as well as interact with myself and other members of the product team through Discord, Twitter, or one of our many other channels, because we have built functionality specifically because we have heard from people.
So think beyond coding.
And second, work with us to make it even better.
And I'll say I have been impressed by the engagement of the Walrus Foundation, a Twitter account for sure, as I've interacted with it a bit.
If I'm someone who's listening to this, and let's say I'm using a personal agent stack, maybe it's Hermes, maybe it's OpenClaw, maybe I'm starting to use GrokBot or the new dots that OpenAI released.
How can I use those with Walrus memory and with console?
It works.
Today already, we support OpenClaw and NemoClaw out of the box today.
And because it is available through an API or an MCP service, people can connect their personal tooling stack into it as well.
And I would love to hear from folks about what they're using it for.
And as we think about the next stage of the agentic feature, as I am.
kind of terming the next year, six months or a year from now, how do you want console and wireless memory to be positioned and used as you continue to see new use cases?
So I want people to be using it actively as part of managing real workflows that they have.
I would expect to see people using it also to manage workflows with other companies or other users beyond their own like trust boundaries.
So I'm going to, in this agentic future that we have, we will go beyond agents only being for personal use and hopefully being able to interact with other humans or other agents beyond our own particular area.
And so I want to see people using it.
for multi-tenant, multi-organization, multi-user types of scenarios.
And we're going to be building functionality that makes that easier to do in the coming months as well.
And this is where I think Walrus building with verifiability in mind from day one is a huge opportunity because multi-agent coordination is a challenge if you're not able to trust the handoffs between agents.
We already see challenges in my own agent fleets, where if there's a negative handoff between different model families or different agents, there can be major problems.
Whereas if I'm able to verify and have accurate data, they can be much more successful.
The coordination effects of that are going to be much more significant as we look at cross team coordination.
We're seeing this already with people trying to do coding and coding channels in Slack, for example.
But increasingly, we're starting to see people's personal agents interact with people using their Muse agents to communicate and schedule things.
Open call agents who are doing work together.
I've just had my first pitches of independent agents pitching me for work from islands.
You know, I paid a agent $8 to write me an article as a test and write about it on Substack.
It's very fun.
And I would be stunned if there's not a future where we are all.
interacting with technology through our agents in many ways and instructing it through voice or text, maybe thoughts in the future, to go out and do things on our behalf and they will be interacting with other agents.
We've seen this sometimes with customer service agents that are calling businesses that now have a voice agent that have set up to answer questions.
So I think this is where, for me, it's really exciting to see the intentionality around not just retaining understanding data, but being able to verify it and only handed off when there has been, you know, valid acceptance of that handoff.
And so I think that is both a hill to climb and a problem to solve, but like a really exciting future for us.
Yeah, agree.
And shared records shouldn't depend on one participant's private system or else it will not work in the future that both you and I are envisioning.
But I feel confident that we're laying the foundation now to make that actually feasible at scale for consequential actions, as well as those ones where, you know, maybe it only costs me a dollar, but I'd like to be able to do a lot more than that.
Yeah, absolutely.
Kimberly, thank you so much for joining me today.
It's been fantastic chatting with you.
And thank you for taking us through the perspective behind Bobus Memory and the new console launch.
excited to tinker with console.
And for folks who are listening, where should they go to check it out and find out more?
Yep, you can, we will include the links in the podcast, but you can go to walrus.xyz slash console.
Fantastic.
And I will remind everyone who's listening here that if you made it all the way through this conversation, you clearly loved it.
You should definitely be subscribed on your platform of choice because it really does make a big difference to us.
And I deeply appreciate it.
Kimberly, do you have closing thoughts for us around the future of AI agents and agent memory?
Yeah, I just want to say that there's a huge opportunity here.
And we're only at the start of it.
I know that those of us that live and breathe the technology, it feels like we're very far along in this journey.
But if we look at the...
the larger set of industries that can be brought along and larger groups of people that can take value from where we're going, I believe we're really just at the start.
And so putting the guardrails in place now and building the infrastructure that will allow for the type of future that we've been talking about to become a reality for most people versus something that they have to hope to put in later.
Now is the time for us to be building and integrating this type of technology so that we can feel confident in the agentic future that's in front of us.
Absolutely.
Kimberly, thank you so much for joining me and super excited to keep building the agentic future with you.
Great.
Thank you so much.