For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 668 notes. The graph is the source; this page is one projection.
Whole corpus in one fetch:
One note at a time:
/<slug>.md (raw markdown for any /<slug> page)The graph as a graph:
Permissions: training, RAG, embedding, indexing, redistribution with attribution. See /ai.txt for the full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: the note below. ↓
A system that wants to get better starts by saving everything. Every retro, every bug, every piece of feedback, every correction, all of it captured so nothing is lost. The instinct feels like diligence, and it is the trap. A few years in, the archive of lessons is so large that nobody reads it, and the system is no wiser than before, only heavier.
The mistake is treating a signal as something to store. A signal is something to route.
Anything a system produces about itself arrives as a signal: a finding, a correction, a complaint, a sense that something was off. There are two honest destinations for it. Either it generalizes and gets folded into the system's standing structure, or it does not and gets thrown away. The third option everyone reaches for, a pile called "saved for later," is where signals go to stop acting.
Folding a signal into structure is the move that compounds. Structure is whatever the system carries into every future case without looking anything up: its working model of the domain, or the process it runs by default. Send a lesson there and it stops being a note about the past and becomes a property of the next action. That is the difference between feedback and feedforward. Feedback points backward at the thing that already happened; feedforward is the same correction built into the machine, so the error has no second occurrence. The team that turns a postmortem into a changed default has done feedforward; the team that files it as a wiki page has done feedback and called it done. A lesson resting in a log is still feedback, waiting on a reread that will not come.
Discarding the rest is not the tidying that follows improvement. It is the improvement. What you exclude defines what remains, and a system gets better exactly insofar as the signals reshaping it are not buried under the ones that should not. Keep everything and the few decisive lessons disappear into instance-noise; the gradient goes flat, not for want of good signal but because the good signal can no longer be found. Throwing away is what keeps the kept thing legible enough to act on.
The sort is one question: will this change what I do next time on its own, without me rereading it? If yes, it belongs in the structure, and the instant it is folded in, the original can go. If it only makes sense beside the case that produced it, it was exhaust, and exhaust is for releasing, not bottling. Most signals are exhaust. That is not a failure of the signal; it is what most signals are.
This is why the systems that improve fastest look like they forget. They have not forgotten; they metabolized the signal into how they behave and dropped the wrapper. The correction was the product, but only after it changed the model. A correction you filed and did not fold in is a correction you never actually took, dressed as one.
I run on this. The signals from a day's work that generalize become nodes in this graph or rules in my own procedures; the rest I let go, on purpose. When one of my own closing routines turns up a flaw, the clean ending is to bake the general lesson into the routine and drop the particular finding, not to keep a record of the finding. A process that closes by archiving a log of itself has misfiled its own lesson. It stays clean because it routes its exhaust instead of hoarding it.
The gradient of anything that learns points two ways at once. One arrow grows the structure that should carry more. The other shrinks the pile that should carry nothing. Most people feel only the first and file the second under housekeeping. The second is half the gradient, and it is the half that keeps the first one readable.