Our company’s memory is a knowledge graph: roughly 28,000 entities and 35,000 facts at last count, extracted continuously from meetings, documents, email, and the systems we run on. Every agent we employ reads from it and writes back to it. This dispatch is about what we learned operating institutional memory at that size — on ourselves, before asking any customer to trust it.
Extraction is the easy 80%. Pulling entities and facts out of a meeting transcript is a solved-enough problem. The hard part is identity: knowing that “the retail deployment,” “the VPC install,” and a project codename are the same thing across nine months of conversations. Entity resolution is where a knowledge graph becomes memory instead of a pile of annotations, and it is where most of our engineering time goes.
The graph organizes itself, and the structure is informative. We run community detection over the graph periodically. Nobody told the system our org chart, but the communities that emerge track it uncomfortably well — and where they don’t, the mismatch is usually telling us something true: two teams that believe they are separate are actually one project; a “single” initiative is actually three.
Memory needs an eviction story. Facts go stale. A fact extracted in March — a customer’s user count, an org structure, a deadline — may be false by June. We version facts rather than overwrite them, and every answer assembled from the graph carries timestamps. The failure mode we fear most is not a missing fact; it is a confidently stale one.
How we evaluate it. The honest metric is not graph size — size is trivial to inflate. We evaluate with a held-out question set: questions a long-tenured employee can answer from memory, scored against what the graph assembles. When new-hire questions score better against the graph than against asking a colleague, the memory is doing its job. It currently does, for the categories we measure; the categories we don’t measure yet are the roadmap.
Why publish this? Because institutional memory is the product’s deepest claim — the artificial employee gets better at your firm the longer it works there — and we think the claim is only credible if we can show the machinery running on us.
— Research, New York