Context and memory

Things about what gets sent up to the model and what stays outside it: context compression, retrieval without vectors, second brains — and arranging a repository so that both people and agents can read it.

That last one looks out of place at first. Karrot SEED is a design system, and what actually came out of it wasn’t a component but the way its documents were laid out. The test for this bucket isn’t what the subject is; it’s whether it’s about deciding what a machine gets to read.

Eleven, and not one line of code crossed over

Subject	What came across	Shape
Headroom	Attach a confidence interval and a label to anything unmeasured	one convention
Karrot SEED	Auto-run the build when the source changes · a verification reminder at session end	two hooks
RAG without a vector DB	Cutting the problem along a scale axis and drawing the boundary	one structure
A second-brain workshop	Score a new skill within the same turn it was written	one timing rule
Graph engineering	Three review lenses in parallel — 9.9 minutes a run	one wiring change
An LLM wiki in Obsidian	One item exactly one place · log model access	two hygiene rules
A second-brain interview	“Context before automation” — many tools, thin criteria	one attitude
A file management system	A home for the unclassifiable · cap the index	two rules, two corrections
A second brain in four steps	Nothing did — instead it found an empty cell of mine	one adoption, one confirmed defect
dsh-anchored-standard	Nothing did — there was no hook to install it into	one rejection, 24 days unrun confirmed
public-apis	Put a cheap check ahead of an expensive one and the expensive one looks like it runs while never running	one trap, one confirmed defect

All rules, hooks, structures and wiring. Zero code ported. Context problems mostly don’t yield to someone else’s code — the answer lives in the shape of your own data, so the only transferable part is the ruler.

And this bucket found the most holes in my own work

Seven pieces used somebody else’s principle as a ruler and produced a number.

• Of 1,021 markdown files in my vault, 47 (4.6%) have front matter — so when retrieval fails there, the cause is likely empty metadata, not a missing tool
• Cache reuse 92.5–96.3%, uncached input 1.8–2.0 tokens per call — the compression tool had nothing left to compress
• Of seven techniques I’d written down as harvested, five died the moment I opened my own settings.json — already there, or a false lead
• A full scan of 99 skills found one with no description at all, so it never fires from natural speech, plus sixteen over 500 lines
• Across 52 run logs, the feature a video called “the heart” met its trigger condition once
• Running the autonomous-accumulation track for real: zero items, zero files, zero observer hooks — it was only installed
• Opening the API catalogue I built a month ago: liveness was measured on 300 of its 1,791 rows (17%)

The third stung most. It was the kind of error no amount of careful reading of the repo would ever catch.

The second axis — what crossed was “when” and “where”

At three pieces only one line showed: no code crosses. Six more, and a more accurate sentence appeared — most of what crossed was not “what” but “when” and “where.”

From the workshop deck what I got wasn’t a new inspection tool but when to run the inspection. From the graph piece it wasn’t a new stage but when to run existing stages together. From the file-management piece it wasn’t a new folder layout but where to put things that don’t classify. All of it moved the timing and the location of something I already had.

I have almost never gained a new part from someone else in this bucket. What I gained was judgement about when and where to use my own parts.

The third axis — the most valuable thing was agreement from different grounds

The file-management piece is the extreme case. That video matched a design I had built separately on five conclusions, and not one of the reasons overlapped — his was “make it complicated and people run away”, mine was “a path that depends on a taxonomy breaks.” Reach the same point down different roads and the odds that the point is right go up.

The reverse also showed up. The same technique was harmless on one side and fatal on the other — numbering folders is fine on ID-based storage and not on path-based storage. Stopping at “it works for them, not for me” would have left a preference; asking why it splits produced a rule.

The fourth axis — “I already have this” is the most dangerous verdict

The ninth piece shook the three axes above. The verdict I reach most often in this bucket is “I’m already doing that” — and I had never once asked what I was reaching it with.

Comparing the autonomous-accumulation step, I wrote equivalent on the grounds that it was installed. Pressing it for real gave zero items, zero files, zero observer hooks. It isn’t that promotion was zero — there was no observation, so there was nothing to promote.

Possession is not operation. And it doesn’t stop at that cell — all four “equivalent” verdicts in that table were judged on installation. The other three are unpressed, and until then they are ledger entries.

This axis meshes directly with the conclusion in agent harnesses. The sentence that came up most often over there was “I already had this” — and the method for checking whether that sentence is true came from here.

The tenth dug the fourth axis one layer deeper — “never switched on”

The ninth was material with almost nothing to teach, which is precisely why I filled the comparison table to the end — and an empty cell fell out of it. Material with nothing to take still earns its keep: when you borrow someone’s ruler, the ruler doesn’t have to be excellent. The tenth proved it again.

If the ninth found possession is not operation, the tenth opened the cell below it. Possession is not a record of running.

That piece went in to tear down someone’s preset for shrinking the context budget. The investigation stopped before any adoption verdict — my harness has no hook that subtracts injection at the first request of a session. With nowhere to install it, the answer is neither adopt nor reject but not applicable.

Something worse surfaced in the same spot. I had brought in a measurement skill half a year ago for exactly this purpose, and it had not run once in the 24 days since it was installed. Zero outputs; the command meant to back it was never even written. And on the same day, the investigation on the harness side pointed at the same skill without knowing about this one.

So the fourth axis grows a floor:

Layer	What to ask
1	Do I have it — is it installed
2	Does it work — press it, does output come out (the ninth)
3	Has it ever been switched on — is there a last-run date (the tenth)

Write “I already have this” off layer 1 alone and you conclude you should go buy the thing that is already in the warehouse. That is the verdict this bucket reaches most often.

One more line. This piece also confirmed that my own setup is itself a wholesale copy of someone else’s preset bundle — over half the skills, most of the rules, ten of the hooks came from there, and exactly two hooks are mine. I went to appraise someone’s preset and confirmed I was already sitting on one. And the ruler I picked up to measure it with was still in its wrapper, 24 days on.