Google's company memory is personal

Google has reached the same conclusion as Snowflake and Databricks: an agent is only as useful as what it knows about your company. I think they’re right, and Google is better placed than either data platform to act on it, because so much of what a company knows passes through Gmail, Chat and Docs.

What Google built

The first piece of Google’s answer is Workspace Intelligence, announced at Cloud Next on April 22. Google describes it as turning “your scattered emails, chats, and files into a cohesive knowledge graph”, one that understands “your active projects, your collaborators, and your organization’s domain knowledge”. It reads from Docs, Slides, Gmail, Sheets, Chat and Drive, and connects to Asana, Jira and Salesforce. It’s organization-wide, and it’s built for retrieval.

The second is Memory, in Gemini Enterprise. Google’s documentation says that as you use the product, “it builds a personal Memory by understanding your individual needs and work patterns.” It learns from your conversation history with Gemini and from connected sources like your mail, calendar and documents. You can also save a fact directly by starting a prompt with “Remember that…”, and remove one by asking it to forget. Admins can turn Memory off for everyone, but it’s on by default.

The third is Gemini Enterprise Projects, a shared knowledge base a team builds around a topic. Members add files from Drive or SharePoint, upload documents and write notes, and linked files stay in sync with their source.

So the organization gets a graph to search, each person gets a memory of their own, and a team gets a knowledge base it fills by hand.

A company decision in one person’s memory

The documentation’s example of a saved memory is this prompt:

“Remember that all new proposals for Project Walby should be reviewed by @Roberto Moreno.”

Gemini replies that it created a new memory, and from then on that person’s Gemini has the decision on hand.

What got saved isn’t a preference, like a writing style or a favorite format. It’s a decision about how a team works, it governs everybody who writes a Walby proposal, and Google chose it to illustrate personal memory.

Imagine someone asks Gemini who reviews Walby proposals. Unless the decision made it into their own Memory, Gemini answers them from retrieval. The graph may find the thread where the team discussed it. It may also find an older thread where somebody named a different reviewer. Both match the question, but nothing in either one says which the team settled on. Or the graph may find nothing relevant at all. So they may get an answer that sounds just as confident as the first person’s but is wrong, or no answer at all.

Now imagine the decision changes. Roberto hands Walby to someone else. The person who saved the original memory can tell their Gemini to forget it, if they remember to. Anyone whose Memory picked up its own copy has to clear it too, and everyone else goes back to searching.

A 2018 survey of 1,001 US adults by Panopto and YouGov found that 42% of institutional knowledge is unique to the individual who holds it. Memory makes that knowledge useful to the person who has it, but keeps it where it already was: with one person.

Search finds what was said

In 2005, at Google, I led the team building a web-scale database of facts about entities, which powered question answering on google.com. A database of facts and an index of documents answer different questions. An index tells you where someone mentioned something. A fact store tells you what’s true, and needs some way to decide between two sources that disagree.

A company needs a fact store too, and keeping one current is harder than on the web. When a public fact changes, like who’s president, news outlets report it everywhere at once. When a company fact changes, like who reviews Walby proposals, someone may mention the change in just one thread that a few people see. A useful company memory has to hold what was decided, who decided it, where that happened, and whether it still holds. Retrieval over messages can find what was said and who said it, but only in messages the person asking is allowed to see, and it can’t reliably tell you whether a decision still holds. I made a longer version of this argument about data warehouses in The context your warehouse can’t reach: one thread floats an answer, a later thread settles on a different one, both match the query, and the easier one to retrieve isn’t reliably the current one.

Of Google’s three pieces, Projects comes closest to a shared company memory, because a whole team can read it. But someone has to decide what goes in and write down the decision. That step depends on someone remembering to do it, which is why wikis rot and why decisions like the Walby one go unrecorded.

What a shared memory would have to do

Here’s how the Walby decision would play out in a shared memory that keeps itself current from the messages where decisions get made. The team makes the decision in a thread. Something reads that thread, notices that it changes who reviews proposals, and proposes an update to a shared page, linked to the message it came from. A person on the team confirms it. When Roberto hands the project off, the next thread proposes the next update, and the old answer is replaced. Everyone’s AI assistant gets the same answer from the shared memory.

Nobody in that loop had to remember to type “Remember that…”. The decision was already in a message.

That’s the layer we’re building at this+that. Our brain is a set of pages holding what a company knows, assembled from the messages it already has across Gmail, Outlook, Slack, Microsoft Teams, and Google Chat, and kept current today by workflows that write back to it as they run, with a person approving or editing each update. We’re building toward a brain that keeps most of itself current from the message stream through one maintenance workflow, and that part isn’t finished.

Run the test on whatever you’re using, ours included. Pick a decision your team changed recently, like who owns a renewal or who signs off on a proposal, and ask your assistant about it from the account of someone who wasn’t in the thread.