The problem with today's memory solutions
A wave of AI memory products has arrived, and almost all of them are storage layers dressed up as cognition. Here is what they get wrong, and why it matters.
Agent memory is having a moment. Zep, Letta, Mem0, Supermemory, and a dozen others have shipped in the last two years, all promising to fix the fact that agents forget. Most of them succeed at making one agent less forgetful. Almost none of them make an organization smarter. The distance between those two things is the whole problem, and it is worth being precise about where it comes from.
They are single-tenant by design
Every one of these products starts from the same premise: memory belongs to a single user. Each agent runs in its own private thread, with its own history, and nothing it learns ever reaches anyone else.
The premise is understandable. It mimics how human memory works, which is the first template anyone reaches for, and the alternative is genuinely hard to build. Organizations are multi-author, hierarchical, and full of opinions that contradict each other and shift over time. A private scratchpad sidesteps all of that.
It also sidesteps the point. A company is not a collection of isolated threads. Give every agent its own private memory and you do not get an intelligent organization; you get a larger number of isolated minds working in parallel, each one rediscovering what someone else already worked out down the hall.
They treat memory as retrieval
The deeper mistake is treating memory as a retrieval problem in the first place. Most of these tools are, underneath, a faster way to fetch the transcript of a past conversation or the chunk of a document that matches a query.
That helps inside one thread. It does very little for the company. A company's knowledge is not a pile of transcripts you can search, or a folder of documents you can rank. It is all of that fused into a working understanding of how the place operates, carried forward by the organization itself. Retrieving yesterday's conversation more accurately never adds up to that understanding, no matter how good the ranking gets.
They are fighting on a closing axis
Look at how the field competes and you see retrieval benchmarks: precision, recall, scores on LoCoMo, how much context fits in a single prompt. The engineering behind those numbers is real. The ground under it is moving.
Models keep getting better. Context windows keep getting longer. Embeddings keep getting cheaper. Retrieval is converging on the point where it is simply good enough, and once it is, the gap between the best pipeline and a merely decent one stops being something anyone outside a benchmark can feel.
A better retrieval algorithm stops being a moat the moment the models catch up.
A company whose advantage is a slightly sharper retrieval pipeline is building on ground that is already eroding.
What the problem actually is
The thing that has to be built is not a better memory for one agent. It is a shared memory for many: one understanding of how a company operates that every agent draws from, gets sharper with every interaction, and carries forward as people come and go. That is a different problem than retrieval, and it is the one that actually makes an organization more capable over time.
It is also the one we are building Glen to solve. If you want the longer argument for why, that is our thesis.