orange tomato
Page 1/10

WHAT THE AGENT REMEMBERS

A companion notebook to the deep dive — notes, synthesis, instructions, and what upkeep earns
NOTES
→
WIKI PAGE
→
STANDING FILE
  • Three places an agent's material persists — and one failure they share: a confident sentence, still cited, no longer true.
  • The question: what should persist for an agent to build on, and what upkeep earns its cost?
410 → 365
One erratum. The wiki remembered the box weight as the lantern. The sentence outlived the shelf.
orange tomato
Page 2/10

1. THE SUBSTRATE — WHAT A COPY KEEPS

  • A note habit produces three things: readable text, followable names and locations, and the person's record of why the note exists. The app adds conventions and a computed layer.
  • Four notes copied and read with plain tools: text held completely. What did not survive the copy — type (values travel as strings, types stay in settings), resolution (links resolve nothing until something looks them up), repair (the cache rebuilds app-side).
  • One link pointed nowhere, visibly identical to the live ones. The Markdown path needed two conventions (relative folder, %20) — the first pass found one cause, the rerun found both.
text
readable anywhere
names
followable paths
context
why it exists
computed layer
stays in the app
Three questions for your own folder: rename survival, one word for one thing, the reason written inside.
orange tomato
Page 3/10

2. THE CACHE — FOUR TASKS, ONE VERDICT EACH

  • Lookup (one-off): tie. Both paths returned the figure; the wiki's 25-minute build cost sat unamortized.
  • Synthesis (repeated): wiki wins. One section answered from prebuilt synthesis what retrieval rebuilt across five opens. Pre-paying repeats pays off — only where the synthesis repeats.
  • Contradiction (dated retest): retrieval stated 24 vs 19h as unresolved; the un-updated wiki said 24h confidently. Fix: ~15 minutes, a conflict note, both citations.
  • Correction (erratum): retrieval preferred 365g and softened the tradeoff (90g → 45g margin). The wiki kept 410g as settled synthesis — the error promoted from spec line into conclusion. Fix: ~5 minutes.
24 vs 19
Both figures, both dates, one unresolved lot. The honest answer names the conflict instead of picking.
Synthetic corpus, manual-keyword retrieval proxy, one operator, informal timings as bounds — the logic of the tradeoff, not production rates.
orange tomato
Page 4/10

3. SIX MONTHS OF LIVING WITH ONE

  • 950+ live notes from ~340 sources. What compounded: context — the system knowing projects and preferences, no re-explaining.
  • 614 notes archived (~40%). Accumulation slowed every addition and raised token cost until the spring clean. What never compounded on its own: maintenance. It waited.
  • ~200 links added late. Extraction ran ahead of weaving — claims captured but never linked or checked — until diagnostics surfaced the backlog. Extracting is step one; connecting and verifying are the rest.
40%
Nearly half the vault turned out to be dead ends, dropped ideas, and detail nobody would revisit. Curate toward thinking, archive the rest.
Single operator, self-reported, plugin-authored rules — a firsthand account with stated limits, converging with the four-task test.
orange tomato
Page 5/10

4. TWENTY QUESTIONS, THREE WAYS

  • 166 business files, 20 real questions (facts, history, decisions, connections, new information), each asked three ways: no files, the author's folder setup, a freshly built wiki.
  • Strict grade: 10 vs 5. Loose grade: 13 vs 10. The folder won both. The wiki cost 154 minutes and $54.39 to build, used 4.6% fewer tokens — a 6,000-question payback.
  • Losses land where the promises run strongest: connections tied-or-lost, new information 3-to-1 against. Per-type: facts 2–1, history 2–1, decisions 3–2.
  • Limits travel with it: one business, bulk build (not supervised one-at-a-time), two grading rounds (direction held), an advantaged setup (18 of 20 answers rest on the author's notes).
0 of 20
With no files at all. Business context is what makes the answers useful — folder first, wiki only for a reason.
orange tomato
Page 6/10

5. THE CONVENTIONS LAYER — ADVICE, THEN EVIDENCE

  • GitHub's 2,500-file analysis describes recurring patterns (commands early, real snippets, tiers) but defines no "successful" — hygiene, not lift. Patterns may reflect authors, not causes.
  • Efficiency study (124 PRs, 10 repos, one agent): with-file runs ~29% faster median, ~17% leaner output tokens. Correctness explicitly out of scope.
  • Success study (4 agents, 300 + 138 tasks): no significant success effect either way (p = 0.87 / 0.37); cost up 20%+; developer-written beats generated by 7%; generated helps (+2.7%) only where docs are absent.
  • The purchasing rule: a file earns its tokens where it carries what the repo does not already say. Everywhere else it duplicates — and duplication is rent.
speed ≠ success
One paper prices time, the other prices solved tasks. Never let one stand in for the other.
orange tomato
Page 7/10

6. THE PATTERN — RENT, PAID THREE WAYS

  • Wiki: minutes per change — the conflict note, the erratum ripple, the 200-link catch-up.
  • Notes: naming agreement — same property names, no dead links, the reason written inside.
  • File: context tokens — every line present on every run, whether or not the task needs it.
  • Redundancy decides: persistence earns its keep where the source is absent (undocumented repo, weekly synthesis) or agreement is load-bearing. Elsewhere, demand a receipt.
receipt
A date on the synthesis. A lint report read, not filed. A file line tied to a failure it fixed.
Surfaces teach each other: the wiki's update discipline ports to notes; legibility ports to the file; failure-first ports to the wiki. Resolve every claim two ways; trust the disagreement.
orange tomato
Page 8/10

7. WHAT YOU CAN INSPECT — THE CHECKLIST

  • Notes: storage page, config boundary, property-type rule, cache paragraph (all re-read Nov 11). Then your folder: rename, same-word search, the missing reason.
  • Wiki: synthesis date on the answering page; the conflict note or its absence; whose lint ran last and what it flagged; the last ingestion bill. Read the cited source, not the sentence.
  • File: what was measured, on which tasks, at what cost — efficiency and success separately. Which lines exist because a failure demonstrated them. What the repo already documents.
Re-check on the surface's schedule — wiki monthly or per batch, folder on decisions, file on failure and after cleanup. Date each check.
orange tomato
Page 9/10

8. WHAT CHANGES THE JUDGMENT

  • Dated as of Nov 11: docs wording current vs a desktop line moved past the check (1.8.9 → 1.15.x); GitHub page live, claims unchanged; overload primary still walled; both file papers call for more agents, models, larger tasks.
  • Open: a published error-rate study for maintained wikis; a file-effect replication across harnesses; the overload primary, inspected; whether the wiki gap closes over months.
  • Until then: persist what gets re-checked, in the surface whose upkeep actually gets paid — and distrust every persistent sentence in proportion to the time since someone verified it.
as of Nov 11
Every observation in this notebook carries its inspection date. Re-check before reusing.
orange tomato
Page 10/10

9. SOURCES & METHOD

  • Docs: Obsidian Help from the official repo source (storage, config, properties, links, attachments, search, backlinks, graph) — current wording, Nov 11.
  • Carried firsthand: four-note copy check (one machine, bounds stated); four-task Bluefin test (synthetic corpus, proxy retrieval, single operator).
  • Others' firsthand: six-month wiki account and 20-question business test — attributed with limits (single operator; one business, bulk build, two grading rounds).
  • Papers: paired-PR efficiency and file-effect benchmark studies, full texts read Nov 11, designs beside every figure. Vendor claims labeled. No model or test execution; the pre-registered comparison stands superseded. Evidence ledger and a non-independent second pass sit in the run record.
Nothing here travels without a source path. Unresolvable claims did not move.