One metaphor, the amnesiac genius, explains the agent's multi-repo and memory problem — then a live demo fixes it without changing a line of code in any repo
A Genius With Amnesia - Victor Savkin, Nx · Victor Savkin
20 min total·Actually worth watching closely: ~9 min·3 must-watch clips
- 0:05 – 1:20Listen
Opening metaphor: the genius with amnesia
A genie grants you the best engineer there is, then imposes restrictions: he can only see one small part of your codebase, maybe one thousandth of it, and remembers nothing he did before. Every conversation starts fresh — and that's what agents are.
A genius on one side and something deeply deficient on the other: the problem isn't intelligence, it's what the agent can see and what it remembers.
Pure spoken argument setting up the metaphor; nothing critical on screen, fine to listen to on a commute▶ Jump to 0:05Speaker · Victor Savkin - 1:20 – 4:50Skim
Two problems: the repo boundary and amnesia
Counts the re-explanations in one change across four repos: publish the UI library, re-explain it in Module 1, go back to UI when it doesn't work, re-explain the fix, do it all again for Module 2, then Platform, then a production bug a week later.
Seven explanations for what is essentially one change — and every box in the diagram can be a different person, so the human ends up being the agent's memory.
The four-repo diagram around 79s is worth a glance; the rest is slide-backed argument you can skim without losing anything▶ Jump to 1:20Speaker · Victor Savkin - 4:50 – 6:39Skim
Your organization's work is a graph
Lays out the full model: a repository graph at the bottom — the artifacts you produce plus tens of thousands of open source repos you depend on — and at the top all the agentic sessions that create and modify that code.
The graph describes what's there and how it came to be; what the agent actually sees is one session, one small fraction of the codebase, no memory.
The graph slide at 289s-310s is the only diagram of the talk's core idea — take in that one image and you can speed through the narration▶ Jump to 4:50Speaker · Victor Savkin - 6:39 – 10:25Watch
Polygraph: the illusion of one big codebase
Demos the agent-agnostic meta harness: analyze the thousands of repos a GitHub user already has access to, extract metadata off to the side, and stitch it into a unified dependency graph.
The fix lives outside the repos — no line of code changes in them, and the metadata layer gives the agent cross-repo vision.
From "let me show you" at 399s this becomes a live demo; strongly visual, and the product interaction only makes sense with the screen in view▶ Jump to 6:39Speaker · Victor Savkin - 10:25 – 13:05Watch
Writing is the hard part: multi-repo PRs and CI
Creates a session, picks repos and an installed agent, states the intent once and lets the change land across both repos; then explains treating all the CI as one vector so a failure resolves to who fixes it.
Reading is the easy half — 10 repos means 10 pull requests and CI to coordinate; the writes are the real engineering problem.
The second "let me show" at 624s opens the most important demo of the talk (one change landing in two repos at 688s) — watch the screen▶ Jump to 10:25Speaker · Victor Savkin - 13:05 – 15:47Watch
Sessions teleport: across people and across agents
With intent, repos, pull requests and traces captured, a session can be reconstructed on any machine, and whoever picks it up can run a different agent — Claude in the original session, Cortex in theirs — while still sharing memory.
Session memory isn't bound to a person, a machine, or one agent: the full state of your session materializes on theirs.
The cross-machine reconstruction at 785s is the most counterintuitive image in the talk — worth seeing for yourself that same repos, same history, different agent really holds▶ Jump to 13:05Speaker · Victor Savkin - 15:47 – 19:59Listen
Referencing a session to fix a bug, and the hive mind
A week-old production bug takes only "this happened, there is a bug" — no extra information required. And instead of selecting repos and sessions by hand, you can ask the graph to find them.
With the space and the time limits both gone, an agent can have more context than any single developer — almost like the Borg, one big hive mind.
The closing section is spoken argument with no strong visual dependency; the reference-a-session bug case at 947s comes across fine by listening▶ Jump to 15:47Speaker · Victor Savkin