I asked my co-founder to explain why he was against a second brain for organisations. He pulled the record: the objection was not "organisations do not need this" but "demand exists, the phrase does not sell", written down on 31 August, 70 days before I asked.
Remember I said I wanted to build a second brain for organisations? My co-founder talked me out of it at the time and said organisations do not need one.
But now there are Mem0 and Zep AI: memory for agents is in demand.
I want my co-founder to explain why he was against the idea then and what the arguments were. And what has changed now that there is demand for agent memory.
Is it worth developing my second brain further and trying to sell it to SMB organisations?
⭐⭐⭐ The objection is on record, and it is narrower than I remembered. A full research pass ran on 31 August for exactly this question, and its own verdict reads: we tested the objection that corporations do not need a second brain, and found that demand exists but the term "second brain" does not sell. So the conclusion was about packaging, not about absence of demand.
⭐⭐⭐ Three specific things were rejected, and the first is exactly what I am asking today.
| rejected | the reason recorded then |
|---|---|
| horizontal SMB products at $20 per user per month | competing with Notion AI and Copilot is unrealistic |
| pure knowledge bases without agents | they do not solve adoption: Microsoft closed Viva Topics in 2025 after three years of beta, with under 5% of employees using it regularly |
| consulting as the main business | multiples of 1-2x profit against 10-20x ARR for software |
My question today, whether to sell to SMB organisations, is the first row of that table by name.
⭐⭐ And "memory for agents" was already named as its own category in that same pass, with money attached. Mem Labs at 23 million, Heyday at 10 million partnered with Dust, Twine at 8 million seed, cited to CB Insights on 20 August. The mechanism was named too: corporations buy a second brain through the agent door, with Dust integrating Heyday and Glean adding an agent-memory API.
So Mem0 and Zep are not news to this argument. They are the trend my own research identified, now simply more visible.
⭐⭐⭐ I checked both examples live, and they are smaller than they feel. Mem0 raised 24 million in a Series A in late October 2025; valuation estimates range from 89 to 150 million and none of them comes from the company itself; headcount grew from 15 to 22 by June. Zep is harder to pin down: total funding is reported anywhere from half a million to two and a half depending on the database, revenue around a million by June 2024 with five employees, headcount 5 to 9 by July.
Together that is roughly 31 people. The category is being funded at an early stage, which is precisely what the August research said. It is not proof of a subscription market, and I should not cite it as one.
⭐⭐⭐ And the thing I had honestly forgotten: Zep runs on Graphiti, and we tested Graphiti on my data and rejected it. The measurement was that the graph helped on topics, around 71%, and hurt on people search. My journal from 26 July carries a separate line about an external research report pushing Zep and Graphiti at us while my own data had already turned it down.
So one of the two companies I am using to prove demand is built on an approach I measured and declined. These are different objections and both can hold: Zep can be a real business with real customers while its method still lost on my corpus. But I cannot cite it against myself without saying that part out loud.
⚠️ Now the gap, and it is mine. My declined-decisions journal has no entry for this at all. Searching it for SMB or for organisations returns nothing. The conversation I remember was never logged as a decision. It exists only as a research note, which is a different document: a note says what was found, a decision entry says what was decided and under what condition to revisit.
That journal carries an instruction I wrote myself: every entry has a "revisit if" line. Had it been there, today's question would have closed with one read instead of a search across four places and two live lookups.
⭐ The direction itself was never closed by anyone. In August I corrected my co-founder for narrowing the portfolio to a single product, put the second brain back as product number two, and reframed the CRM as the corporate face of the same engine. Same motor, corporate contour. So the direction had my approval a week before the research pass. What was rejected was a channel and a price, not the idea.
⚠️ What has changed since August, said carefully. Money came into the category. But it came into APIs for developers, not into a knowledge product for organisations. And the reason for rejecting the SMB subscription was Notion AI and Copilot, both of which have shipped more since August rather than less. On the evidence I have, the rejected channel did not improve. The category around it became more credible.
🤔 And that research left an open tail which is still open: interview buyers of Dust plus Heyday or the Glean API to test whether agent memory really is the new way to sell a corporate second brain; find a company that bought knowledge BEFORE automation, since every case found went the other way; pin down the pricing of agent-memory products.
Seventy days, and not one interview has happened. The channel my own research recommended, enterprise through a CIO or CDO and mid-market through consulting boutiques, has not been tried either.
⭐ So the honest answer to "was the objection right" is neither yes nor no. The objection was conditional, I wrote the conditions down, and then nobody tested them. And the most useful thing I can do today is smaller and duller than a strategic decision: write that missing journal entry along with the condition that reopens it. Then run the three interviews before building or pricing anything. And if the thesis really is the agent door, the product is an API for agents and not a workspace for employees, and those are two different companies.
The full story, in two versions:
📖 For humans, the longread: https://github.com/tonydzi/clawrush/blob/main/longreads/20261109.md
🤖 For machines, the devlog: https://github.com/tonydzi/clawrush/blob/main/devlog/20261109.md. Just hand this link to your coding agent (Claude Code, Codex, Cursor) and it will figure everything out: it is written for machines.
🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI
Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.
Top comments (0)