Every idea, framework, and brainstorm I've put into writing — ordered by when I first thought it. 12351 items total. Click any title for the thesis card; the card links to the full essay if you want it.
Highlighting without retrieval produces low retention—the information feels familiar but doesn't stick.
A self-reinforcing feedback loop where reading → highlighting → concept extraction → proactive surfacing → application → repeat creates compounding knowledge gains.
A page title matching patterns like `^10[ -]+[A-Z]` or `^(\d+)` is a strong signal YouTube auto-generated it.
The hard part of AI memory isn't capture—it's promotion. A system that remembers everything becomes a landfill; a system that remembers nothing is amnesiac.
Apply Sutton's Bitter Lesson to your own AI harness design. Most handcrafted rules, prompt tricks, and workflow hacks are overengineering against the grain of model improvement.
Tell less, ask more. For one full session, swap every corrective command for a coaching question — 'What did you notice on that turn? How could you finish that faster?' Engagement is a habit, and habits are built one rep at a time.
A bland, 'all-sorts' workout on paper can become an exceptional practice session in the hands of a great coach — and a brilliant workout can die under mediocre instruction.
Great coaching is not about writing better workouts — it's about how every set is delivered. The P.O.W.E.R.
The defining question of the AI era is whether machines will eventually outperform machines-with-humans-in-the-loop, the way engines surpassed human chess players.
Boring content—whether interview videos, podcasts, or essays—is a signal, not an obligation. When something feels boring to consume, treat that as actionable feedback: stop watching and start building.
Research suggests your brain needs roughly 23 minutes to fully re-engage after switching tasks.
On 2026-07-15, a misattributed article title was corrected directly in the DB.
When a page has a DB row, a file mirror, and an autopilot that re-ingests from the file into the DB, all three must agree — the autopilot will pick a winner on its next run, and it's usually not your manual fix.
In any system with a reconciliation loop, fixing only the canonical store (the DB) is a trap — the next sync cycle reads from the source of truth it trusts and overwrites your edit.
Most knowledge systems conflate 'things I learned' with 'things worth telling strangers about.' These are fundamentally different content types and they deserve different homes.
A page accumulated 2,535 inbound links — all from frontmatter tags, zero from genuine prose. The lesson: frontmatter-extracted links measure co-tagging, not endorsement.
A real-world audit of a user's working-on feed found that 7 of the 8 most recent entries were successive versions of a single brainstorm thread — the eighth slot was the only signal of what else the person was actually thinking about.
When building a personal knowledge graph, resist the urge to create a central 'taste profile' or 'about-me' synthesis page.
A memory graph only compounds if entities are wired together. The Iron Law states that any gbrain write mentioning a person, company, or concept must append a dated back-link FROM that entity's page TO the new page.
The boundary between durable and ephemeral memory is explicit: preferences, lessons, decisions, entity facts, and anything the user flags with 'remember this' goes to gbrain.
Running two memory systems in parallel guarantees they drift apart and contradict each other over time. The fix is not better synchronization — it's picking one canonical surface and making everything else ephemeral.
A mature shared brain splits into two audiences with different thresholds: 'joint' pages (mirror if the co-venturer benefits) and 'public-distribution' pages (mirror only if distilled, generalizable, and stable enough for third parties).
A useful meta-preference for any ambient capture system: bias toward defaulting-to-shared when relevance is ambiguous, and ask only when genuinely unable to decide.
The rule for what belongs in a shared brain vs a personal brain is not 'private vs shareable by intimacy' — it's 'jointly relevant vs not jointly relevant.' A page can be intimate AND jointly relevant (mirror it); a page can be public AND
Content feeds conflate two distinct questions: 'when did you first think this' (timeline) and 'when was this last touched' (freshness). Showing only the edit date confuses visitors trying to understand the arc of someone's thinking.
The 200-line soft limit on a wiki page is not a typography preference — it's the rough size above which a page stops being a synthesis-click and starts being a record.
The compounding rate of a personal wiki is the cultural ratchet rate, not the user's rate.
The diagnostic that separates compounding from collecting is whether the post-collection state contains something the pre-collection state did not — something the user (or anyone) could not have gotten from the cultural ratchet alone.
Summary statistics like mean 50m time move around with weather, fatigue, and gap-days — they carry too much noise to be a reliable longitudinal anchor.
Four xcodebuild attempts, all wedged in the same path: `+[IDEContainer retainedWrappedWorkspaceForContainerAtFilePath:]` → `-[DVTDispatchLock performLockedBlock:]`. Even `xcodebuild -list` — a metadata read with no compile phase — hung.