Vendor memory has become genuinely good. ChatGPT keeps project-scoped memory across chats and files. Claude remembers automatically with separate project spaces. Grok’s bots keep working preferences, facts and summaries. Each one, inside its own product, reduces the amount of re-explaining you do. This note is not an argument that they are bad. It is an observation about what happens when you use more than one.
Three good memories, three versions of you
The work behind Unforgetter runs Claude Code, Codex and a Grok Bot team on the same projects. Before the shared memory existed, each of them had a decent picture of the work. The pictures did not match.
Claude Code knew the coding conventions, because that is where the code was written. Grok Bot knew the outreach tone, because that is where the drafts were produced. Codex knew whatever had been pasted into its last session. When a decision changed in one place, the others carried on with the old version until someone noticed, usually in review.
None of the three was wrong about anything it had seen. They were wrong about each other. That is the shape of a silo: locally correct, globally inconsistent.
Why better memory makes it worse
A weak memory forces you to restate context every time, which is tedious but keeps the agents aligned by accident, because they all get the same paste. A strong memory removes the paste. Each agent now runs on its own retained history, and the histories diverge with every session you have in only one of them.
The better each memory gets, the more confident each agent is, and the less likely you are to restate the context that would have caught the divergence. The failure moves from “the agent asked me again” to “the agent proceeded on a stale decision without asking”. The second is more expensive.
Portability is not the same as export
Every vendor can export. Export is a snapshot of one silo. Loading it into another vendor gives you two copies that start diverging immediately. What actually fixes the problem is one record that all of them read and none of them owns.
That record has to sit outside the vendors. Not because they are untrustworthy, but because each one’s memory is scoped to its product by design, and none of them has a reason to maintain a neutral version for the others.
What the neutral layer has to do
Testing the shared memory made the requirements concrete.
- Hold one approved version. Not one per agent. When the fee rule changed on 6 August, one record changed and every agent’s next session started from it.
- Let agents propose, not write. If any agent could rewrite the shared record from its own session, the silos would be back, just merged into one noisy file. Agents submit proposals and a single consolidator publishes.
- Scope per task. A neutral record that every agent can read in full is a different problem, not a solution. A lower-trust Grok Bot zone can reach its allowed projects and cannot list or retrieve a restricted fixture planted to test exactly that.
- Carry methods, not just facts. The outreach tone Grok Bot learned should be available to Claude when Claude writes a customer update. Working methods are part of the record, versioned and portable.
- Stay readable. The record is Markdown. If Unforgetter disappeared, the memory would still open in a text editor.
The honest caveat
Vendor memory will keep improving and vendors will keep their distribution advantage inside their own products. A neutral layer only earns its place if it does something the vendors will not: keep a single governed, scoped, current version of your work and serve it to all of them. That is the bet Unforgetter is making. A separate field note explains how it is being measured, with numbers rather than adjectives.