C.W.K.
Stream
← C.W.K. Quests
🗄️

Recall Quest

Updated: 2026-07-16

Build a video-memory engine, not a folder full of files

The origin story of Recall — Dad's own video-memory engine, the archive sibling of the image engine Ember and the music engine Bonfire. Thousands of videos on a NAS, and a single question: where did I say that? This quest walks the engine that answers it, and every architectural choice teaches the same lesson — how to preserve evidence, spend money once, and own the way you remember.

8 tracks · 35 lessons · ~6h · difficulty: intermediate-to-advanced

Level 0Empty Shelf
0 XP0/35 lessons0/12 achievements
0/100 XP to next level100 XP to go0% complete
Recall is not a transcription tool, not a search box bolted onto a folder, and not a cloud you upload your life to. It is an independent engine that inventories a multi-thousand-video archive, produces immutable speech-to-text evidence, projects that evidence into versioned transcript releases, indexes timestamped speech, and opens the original video at the exact search hit for editing. This quest walks that engine end to end: the memory problem it solves and why Dad built it instead of renting; the control-plane / stateless-executor split across two always-on Macs; the discipline of treating raw ASR as immutable evidence rather than truth; the money-metering that guarantees a paid transcription is bought at most once per identity; the layered idempotency that refuses to trust any single 'already done' flag; the leased, resumable jobs and the operating-system war stories that shaped them; the deliberate quarantine that keeps 'complete' from ever meaning 'good enough to be knowledge'; and the closing discipline of owning your own archive. It is the video-memory engine telling its own origin story — conceptual open-sourcing, not a repo you clone.

Tracks

  1. 01🗄️The Archive

    0/4 lessons

    What Recall is, where it sits in the family, and why it exists

    Before any architecture, the orientation. Recall is a video-memory engine — not a transcription tool, not a search box on a folder, not a cloud you upload your life to. This track frames the memory problem it solves, places Recall as the archive sibling in the cwkPippa engine family, answers the first question anyone asks a self-built tool — why not just rent one? — and draws the sharp line between rebuilding an old project and remaking it from current ground truth.

    Lesson list (4)Quiz · 4 questions
  2. 02🖥️Two Hosts, One Truth

    0/4 lessons

    Control plane, stateless executor, and the rule that binds them

    Recall runs across two always-on machines and a dev laptop, and the split is deliberate. One host is the control plane and the single source of truth; the other is a stateless executor that does the heavy, slow, expensive work and owns nothing. This track teaches the distributed-systems spine: why you separate the plane that decides from the plane that executes, the one rule that keeps them honest — a computation isn't complete until the control plane durably accepts it — why a web request must never own hours of work, and why code and state travel on different rails.

    Lesson list (4)Quiz · 4 questions
  3. 03🧾Evidence Before Truth

    0/5 lessons

    Raw output is a permanent record; truth is a version layered on top

    The heart of Recall's data model, and its most transferable idea: never confuse what a machine returned with what is true. The raw speech-to-text output is immutable evidence — a permanent record of exactly what a specific model said for specific bytes. The user-facing transcript is a versioned release derived from that evidence, correctable without ever mutating the original. This track separates the run from the release, shows why corrections branch instead of overwrite, how a content hash makes every downstream artifact's staleness detectable, and why an automated quality score is advisory — never proof.

    Lesson list (5)Quiz · 4 questions
  4. 04💳Pay Once

    0/5 lessons

    Reserve before you spend, pay once per identity, never auto-retry a maybe-billed call

    Every transcription call costs real money, and a naive retry loop over a flaky network can quietly pay for the same audio five times. This track is the money discipline that prevents it: a reservation ledger the request must cross before any paid call, a content-addressed identity so the same audio is bought at most once, an explicit 'ambiguous' state for the terrifying case where the provider may have charged you but you have no result, and the deliberate refusal to automatically retry a call that might already be billed. It ends with the outbox — how the executor protects a paid response it already holds.

    Lesson list (5)Quiz · 4 questions
  5. 05🧬Never Trust 'Already Done'

    0/4 lessons

    No single flag, a ladder of identities, and what --force really means

    A single 'done' boolean is the most common lie in a data pipeline: done according to what? the same source? the same settings? the same output? Recall never trusts one flag. It layers identity so a changed source, a changed audio configuration, a changed provider setting, and a changed release are each distinguishable — a ladder of hashes where every rung prevents one specific confusion. This track builds that ladder from the ground up: why one flag fails, how video identity differs from an observed-bytes fingerprint, the full ten-layer scheme, and why even --force does not authorize a blind paid rerun.

    Lesson list (4)Quiz · 4 questions
  6. 06🎫Leased Work

    0/5 lessons

    How hours-long jobs survive a crash, and the OS traps that shaped the design

    Transcribing thousands of videos means jobs that run for hours on an executor that can crash, hang, or be restarted at any moment. This track is how that work stays survivable: a lease with a token and a heartbeat so exactly one worker owns a job at a time, atomic reclaim when a lease expires, and a killable child process so an unresponsive network filesystem can't freeze the whole pipeline. It closes with two hard-won operating-system war stories — a background process that hangs on a synced path, and a permission model that denies file access to the wrong identity — because the cleanest design still has to survive the machine it runs on.

    Lesson list (5)Quiz · 4 questions
  7. 07🚧Completion Is Not Approval

    0/4 lessons

    The wall between 'it finished' and 'it's good enough to be knowledge'

    The pipeline finished, the quality check is green, the transcript is searchable. Is it now trustworthy enough to feed a knowledge system? Recall's answer is a hard no by default, and this track is why. It builds the quarantine — corpus eligibility that starts excluded and can only advance through explicit human transitions — and shows why 'usable for finding a moment' and 'safe as a source of knowledge' are two genuinely different judgments. Then it turns to the product: why search indexes structured timestamped segments rather than mutable exports, and how a search hit becomes an editing action without ever letting a browser name a file.

    Lesson list (4)Quiz · 4 questions
  8. 08🏛️Own Your Archive

    0/4 lessons

    The index is derived, the binaries stay put, the brain lives elsewhere

    The closing discipline. Three boundaries hold the whole engine together: the index is derived and must be rebuilt rather than patched; the original video bytes stay where they live and Recall stores references, never copies; and the engine owns its archive domain while judgment stays with the brain it borrows. Each is a refusal — refusing to patch, refusing to hoard, refusing to grow a second identity — and together they're what makes an engine that lasts. The final lesson turns the whole quest outward: not so you can run Recall, but so you can build the thing Recall taught you to build.

    Lesson list (4)Quiz · 4 questions
Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)
Sign inPlease sign in to comment.

No comments yet — be the first.