"I pointed Claude at every pull request"
On Organizational Memory
One of the most interesting and under-explored ways AI will enable the future of work is through shaping and delivering organizational memory. I’ve written previously about the beauty of tacit knowledge, arguing that vertical AI solutions play a critical role training and coaching the next generation of workers in high-context industries. Organizational memory powers this process.
It starts with the abundance of data. Platforms can now absorb everything a company produces: text, image, voice, and formal artifacts across existing CRM entries, documents, transcripts, SOPs, and more. But also by capturing every piece of internal data of how a team collaborates together from job descriptions to conversational chat history to actual outputs. It gathers information and reindexes memory as it’s appropriate for the next use case, turning unwritten rules into something that is explicit and portable.
Here’s a real world example. “I pointed Claude at every pull request and comment this year, and made it create a guide on how we review PRs and our code style” says Avi, a friend leading AI at an Series A startup. He uses AI to codify and document engineering team norms:
A team culture is having a sense of our quality versus speed barometer. Part of it is knowing when or not we feel comfortable. When are we okay skipping tests? All these kinds of things that become kind of embedded in the culture, but they’re not necessarily explicit anywhere.
As you can imagine, this becomes a bedrock for automation.
Teams sharing memory across coding tools, for example, can result in multi-factor productivity gains. Fleets of agents could remain consistent to an organization’s institutional playbooks and style while reading and writing from the same memory bank. I imagine the future of memory as model agnostic, where frontier models can be swapped for one or another, but you never want to wipe the accumulated memory of an agent.
What does organizational memory allow us to do?
Create data gravity for customers on new platforms
Memory is a moat. For example, healthcare organizations cling to legacy EHRs for practical reasons, including data gravity across patient records, integrations, and compliance history. Memory creates something even stickier when companies can preserve why past decisions were made (e.g., interpretations, exceptions) and span across multiple users. Collecting all the data upstream of a CRM or store central files, for example, allows you to use the data across files to move naturally into downstream workflows. The challenge becomes transitioning from one system of record to another.
Deploy AI agents that “feel like they have tenure”
Another frame is personalization. For example, in the customer service world, Sierra’s CX agents recall prior conversations, and resolve issues with the continuity of someone who’s been on the account for years. When a platform sits at the intersection of multiple parties: customer and agent, law firm and client, vendor and buyer, it accumulates context that no single party holds. This positional advantage lets AI become an arbiter of quality across a network - building the kind of situational understanding similar to that of a seasoned employee.
Expand the scope of work or client base
Organizational memory can fundamentally change the economics of service delivery. In any client services industry, storing rich customer profiles allows you to anticipate what products clients need next and underwrite that risk in adjacent areas. For example, primary care physicians that can understand the right mix of add on services across conversations with patients. Importantly, owning downstream automation lets you take on clients who would have been economically unviable due to manual overhead. This expands the TAM of markets bottlenecked by expert labor shortages.
The best founders in vertical AI understand that they are the stewards of organizational memory, and now view it as a core competency.
—
There are many more rabbitholes under the broader theme of memory that I will write about soon - including new techniques that make parallels to human memory mechanisms, skills (procedural memory), and more. Stay tuned! Also, feel free to email me through substack, or DM through twitter to chat more about memory.
—
Thanks for joining me today on technopoetic. My posts are intended to be short, living drafts. Look forward to more writing in 2025.
In the meantime, you might like these related posts:
