Published July 16, 2026
A few days ago we wrote about the moment a second AI vendor connected to the same local memory: OpenAI’s Codex, reading and writing the same brain as Anthropic’s Claude. Now a third has joined. DeepSeek is also working with EIDARA.
So the picture today is three frontier models, from three different companies — Claude, ChatGPT and DeepSeek — reading the same compiled memory and contributing to it. About 630 notes, compiled into a single file of roughly 18,500 tokens that each assistant reads at the start of a session.
But the interesting part of this post isn’t the third connection. It’s what happens when you ask three independent models, trained by three different labs, to look at the same system and tell you what’s wrong with it — separately, without talking to each other.
They don’t audit the same way, and that’s the whole value.
Claude brings continuity. It’s the model this system was built with over the past year — it knows where the wires are, which design choices are deliberate and which are scar tissue. Its job now isn’t to invent new ideas. It’s to process the other two’s proposals and decide what actually ships.
ChatGPT brought a security lens. Its sharpest contribution was about permissions — and it was the only one of the three to attack the point head-on. The others assumed that if a client’s access is written down in a config, that’s the boundary. ChatGPT’s objection was blunter: I can’t verify that boundary from where I’m standing, so it isn’t one. A written rule the client can’t check is a promise, not a wall. That single observation turned into a batch of work on how the memory server proves who a client is before it trusts it.
DeepSeek brought two things nobody else did. First, efficiency: concrete proposals to do routine work with a lighter local model and cache repeated decisions, which cut the routine model calls by roughly 40–50% with no measurable loss in quality. Second — and this is the one I keep thinking about — it was the only model that treated the local curator as a single point of failure. The others assumed it would keep running. DeepSeek assumed it would fall over one day and asked what the plan was. There wasn’t one. There is now.
Here’s the thing about a single reviewer, even a very good one: you can never fully separate its insight from its blind spots. A model trained one way tends to notice one kind of problem and miss another.
But when three models from three different labs, reading the same system on their own, independently land on the same diagnosis — that’s about as close to an objective signal as you get. It’s not one house’s opinion anymore. It’s a problem real enough that three different kinds of intelligence all tripped over it.
And they did converge, more than I expected. All three, without coordinating, flagged that the system’s watchdog should act on problems rather than just report them. All three flagged that the memory it hands an assistant at the start of a session had grown too heavy, with too much useful history sitting in a “cold” tier the assistant never saw. All three flagged that the protocol for how models write into the memory needed formal versioning, so a mismatched client can’t corrupt anything.
Those convergences went to the top of the list. Where only one model saw something, we treated it as a lead worth chasing — the ChatGPT permission catch and the DeepSeek resilience catch both became real work precisely because they were the blind spots of the other two.
Cross-referencing three audit lists and building from the overlap, we closed close to thirty structural changes in the first week of the trio. Not new features — refinements. That distinction is the whole point of this phase.
A few, in plain terms: the memory a session starts with is now delivered in layers, so an assistant gets a light overview first and pulls detail on demand instead of swallowing everything up front. The background curator now pushes back when it’s overloaded instead of silently falling behind. The watchdog opens its own tickets when it finds a problem. And the write protocol between a big model and the memory is now versioned, with conformance checks that block a client whose version doesn’t match. In the observation window, with three models writing to the same brain, we’ve logged zero cross-vendor contradictions — the write rules are holding in practice.
Two honesty notes, because they matter for the phase we’re in. The permission fix ChatGPT prompted shipped in a pragmatic form — the server now ties a client to the actual process behind it, rather than trusting a name — but the stronger, cryptographic version is documented as still to build. Showing that gap openly is worth more than a checkmark that isn’t fully earned. And the maintenance backlog those changes were meant to drain isn’t flat yet; it dropped below target for days, then spiked again under a wave of new input. Calling it “stable” today would be inflating the result. It’s better, not done.
One number, framed honestly: on a fixed set of 31 hard retrieval questions we run continuously as an internal check, top-answer accuracy went from 6.5% to 25.8% over this window — nearly four times the baseline. That’s our own benchmark, not an industry one, which brings me to the point of the whole post.
For most of the past year the question was what should this system do? With three models’ input processed and the list drained, the question changes to how good is it, really, next to everyone else?
So the next phase is benchmarks — the public ones the field uses: LoCoMo, LongMemEval, BEAM. Until now EIDARA has been optimized for the real life of one specific user. Benchmarks are the cold comparison against the other systems in this space — Mem0, Zep, Letta, Hindsight — under conditions we didn’t design.
I want to be plain about where we are: EIDARA has no public benchmark score today. Zero. The honest first step isn’t a headline number — it’s publishing our internal benchmark with the full method and the exact questions, so anyone can see how we measure before we claim anything. Then we run the public ones and publish what comes back, good or bad. We asked all three models for a second list, this one only about benchmarking, and we’ll cross-reference those the same way — to separate what’s genuinely worth measuring from each lab’s house style.
The goal was never to win a ranking. It’s to have honest, comparable numbers before EIDARA goes further out into the world.
A year ago, in a post about sharing agents, we wrote: any AI that reads it becomes that agent — Claude, GPT, DeepSeek, doesn’t matter. That was a line about architecture. Today it’s a description of an ordinary week: three models from three vendors, reading and improving the same brain, none of them owning it.
The vendor-agnostic thesis is settled. What’s left is to measure the thing honestly and let the numbers speak. We are still working on this. Every week it’s a little closer to the thing I want it to be.
— Javier
EIDARA v2 is free. SUPER DARA is what comes next.