Personal knowledge systems · Field comparison · Researched 21–22 Nov 2026

Three workflows, three hundred mirrors

Your Dream Cycle, Distillation Pipeline and Capture Loop — each held up against 100 published alternatives, to steal what works. Every entry below was verified from a published source; nothing was invented to fill the count.

300verified entries — 100 per workflow, all three tracks completed
~284independent systems — 16 entries are the same system under another name or track
13entries flagged “thin” — existence verified, detail only mention-level in the source
48web searches behind the comparison, read 21–22 Nov 2026

The three workflows being compared

As you run them today — the baseline every alternative is measured against.

Track 1 · Memory audit

The Dream Cycle

Every night, an automated pass reads the day’s dated markdown log of timestamped claims, plus the running memory file and the open human-verdict queue. Findings are posted as verdicts; you dispose of about 10 a day, and accepted corrections are written back to memory with a line stating what they replace. Nothing is deleted — clearing means moving to Completed.

  • Check 0 — verdicts the day’s evidence already settled
  • Check 1 — contradictions between claims
  • Check 2 — staleness: claims superseded by later events
  • Check 3 — wrong numbers
  • Check 4 — orphans: claims with no corroboration
  • Check 5 — duplicates
Measured: ~5.8 new verdicts/day · backlog ~210 · ~91% of staleness verdicts supersede same-day

Track 2 · Staged refinement

The Distillation Pipeline

Raw captures — audio transcripts, photos, screenshots, documents, voice notes — land in a bronze inbox and are refined through four levels. A ledger tracks every capture so nothing sits unprocessed unseen. The design direction: the ledger lives in the same vector/graph store as the content, levels are computed from artifacts on disk, and the sweep runs as a database query.

  • Bronze — raw capture, as-is
  • Silver — transcribed, summarised, normalised document
  • Gold — synthesised index or guide joining many silvers
  • Platinum — atomic verified facts, each ratified by a human verdict
Four levels — where almost everyone else in the 300 stops at three

Track 3 · Capture to context

The Capture Loop

Continuous, low-friction capture from daily life — wearable always-on audio with transcripts, phone photos, screenshots, voice notes, messages — into one inbox. Intake classifies each item (receipts, job sites, family, documents, whiteboards, screenshots) and files a shortcut into the right folder. Everything is indexed into a vector + full-text + graph store as it happens — continuous indexing, never batch catch-ups.

  • Doctrine: anything not filed and indexed effectively does not exist
  • Trigger: continuous — indexing happens at capture time, not in a later sweep
Always-on capture + continuous indexing + the “doesn’t exist” doctrine — a combination no other entry shares

Headline findings

The nightly audit loop has no close parallel

Across all 300 entries, no other system runs a nightly automated audit over personal memory with specific truth checks feeding a human verdict queue. AI memory systems extract and dedupe; CRM hygiene dedupes enterprise records; spaced repetition audits recall, not truth. The combination is, as far as this comparison can establish, unique.

Platinum is the innovation nobody else has

Staged raw → refined pipelines are the single most common pattern in the 300 — three to five levels with quality gates is mainstream, and your first three levels match it. The fourth level — human-ratified atomic facts — appears nowhere else. The closest parallels are GRADE certainty ratings and Data Vault’s business vault, neither of which is per-fact verification in a personal system.

Vendor mortality is the most consistent failure mode

Pocket, Omnivore, Moves, Mendeley, Limitless and Narrative Clip all died or were absorbed. The lesson repeats across all three tracks: the archive must be local-first, open-format and exportable, or it dies with the tool.

The gaps are about judgement, not mechanics

No prioritisation (a ~210-verdict backlog processed in arrival order), binary ratification with no confidence grading, no stated export guarantee, synthesis that can smooth over disagreement, and throughput metrics with no retrieval-quality measurement. Detail in the five gaps below.

Explore all 300 alternatives

Search by name, what it does, or the idea worth stealing. Open any entry for its stages, trigger, human effort, failure modes and source.

The 10 patterns almost everyone shares

Across all 300 entries — the common ground your workflows already stand on, and where they diverge.

1. Staged refinement: raw → cleaned → synthesised → verified

Medallion, dbt staging → marts, Zettelkasten fleeting → permanent, CODE, Progressive Summarization, PRISMA, OAIS — nearly every durable system converges on 3–5 stages with a quality gate between them. Your four levels are a mainstream instance; the fourth level is the unusual part.

2. Inbox + periodic review: capture now, process later

GTD, BASB, Inbox Zero, Bullet Journal migration, CRM dedup runs. Capture must be frictionless and immediate; processing can be deferred — but only if a scheduled review actually happens. Systems fail when the review is skipped, not when capture is imperfect.

3. Humans decide, machines do the mechanics

Machines transcribe, sync, dedupe, index and rank; humans judge truth, importance and belonging. Systems fail when they automate the judgement or manualise the mechanics (manual start/stop logging, manual entry).

4. Indexing is the product; capture is the easy part

MyLifeBits’ founding lesson: everything captured, nothing findable. Search-first tools exist because retrieval is the hard problem. Systems that don’t index are repeatedly described as “just storage.”

5. Append-only history; nothing is deleted

Delta Lake time travel, Iceberg snapshots, Data Vault insert-only, Kafka’s replayable log — and the Dream Cycle’s “clearing = move to Completed.” Corrections create new versions, so “what did we believe on date X?” stays answerable.

6. Metadata at capture time — context is unrecoverable later

Day One’s auto-attached location and weather, Zotero’s metadata extraction, photo EXIF, Readwise’s “tag while the context is still fresh.” Deferred tagging fails; your intake classifier is an automated version of this pattern.

7. Deduplication as a first-class stage

dbt silver, PRISMA, Rayyan, CRM hygiene, Data Vault hash-diff — dedup has its own stage, key and metrics in mature systems. What’s unusual in yours is running it nightly over personal claims rather than periodically over enterprise records.

8. Quality gates between stages

dbt tests, pipeline expectations, DocETL validation, GRADE, PRISMA inclusion criteria. A record that fails a check doesn’t advance. Platinum ratification is one such gate; the pattern suggests gates at silver → gold too.

9. The ledger, audit trail and provenance

NiFi provenance, PRISMA flow counts, Data Vault record sources, Delta Lake’s transaction log. Every serious pipeline can answer: where did this come from, what happened to it, what’s still unprocessed?

10. Vendor mortality: the archive outlives the tool

The single most consistent failure mode in the 300. The counter-pattern is just as consistent: open-source, self-hosted, local-first tools with export as a first-class feature — Zotero, Immich, Karakeep, Wallabag, Logseq, Obsidian.

Top 10 ideas worth stealing

Ranked by expected payoff for a single-person system with a human in the loop.

  1. PRISMA flow diagram as the ledger’s visualisationAn auditable count at every stage — bronze → silver → gold → platinum — with a reason for every capture that doesn’t advance. Makes the ~210 backlog visible and diagnosable in 10 seconds instead of a single number. Cost: one query + one view. (Track 2, #79)
  2. Demand-driven gold: distil only on revisit or useSweep everything to silver (cheap, automatic), but promote to gold only when a silver is used, searched or cited — or when enough silvers on a topic accumulate. Eliminates synthesis work on material nobody ever needs. (Track 2, #69)
  3. GraphRAG community summaries as the gold algorithmChunk → extract entities and relationships → detect communities → summarise each, lower levels feeding higher ones. Gold guides derived from the graph rather than hand-written, and able to answer global questions naive retrieval can’t. (Track 2, #51)
  4. Auto-enrichment at capture + tag while context is freshAttach all available context at intake — location, calendar event, app open, people present — because it is unrecoverable an hour later. Every downstream step gets better for free. (Track 2 #98, Track 3 #68)
  5. Time travel: “what did we believe on date X?”Immutable snapshots plus a transaction log, with a first-class “as of” date on retrieval. The staleness check becomes a change-feed consumer instead of a re-scan, and disputes become answerable. (Track 2, #36–37)
  6. Train-by-moving: every re-file is training dataWhen you move a capture or change a verdict’s outcome, that correction is the training signal — no separate labelling step. Classifier accuracy compounds with zero extra effort. (Track 3, #63 and #66)
  7. GRADE-like certainty rating per platinum factReplace binary ratified / unratified with graded certainty — high, moderate, low, very low — from source quality, corroboration, recency and consistency. Low-certainty facts get re-verified first; retrieval can surface confidence. (Track 2, #86)
  8. Diarisation + entity extraction in the transcription callSpeaker labels, entities, PII redaction and chapters in the same call as transcription, not as downstream steps. Silver transcripts arrive with “who said what” attached and PII redacted at the earliest point. (Track 2, #62–63)
  9. “Ask your past”: question-based retrievalNatural-language questions over the lifelog, with answers that cite their source captures — a question box, not a search box, as the primary interface. (Track 3, #13)
  10. Tiered processing: Baseline / Moderate / IntensiveAssign a tier at intake and process to that standard, with a defined “golden minimum” for baseline. Effort proportional to value; perfectionism stops creating backlogs. (Track 2, #95–96)

What your three workflows do that almost nobody else does

1. A nightly audit loop over personal memory

Five-plus specific checks, automated detection, a human verdict queue, and write-back with replacement lines. No close parallel in the 300.

2. Platinum: a fourth level of ratified atomic facts

Every staged pipeline in the 300 ends at three levels. Separating “synthesised and useful” (gold) from “verified and true” (platinum) is a genuine innovation.

3. The ledger in the same store as the content

Levels computed from artifacts on disk; the sweep as a database query. Most systems keep a separate orchestration layer or tracking database. Derived state eliminates a whole class of sync bugs.

4. Continuous indexing with a “doesn’t exist” doctrine

Always-on capture + continuous indexing + unindexed captures treated as visible failures. Streaming systems share the mechanics; no personal capture system in the 300 shares the doctrine.

5. One verdict queue across all three workflows

Audit and distillation share a single human-decision surface: one queue, one action (accept / reject), measured throughput of ~10 disposed against ~5.8 new per day.

6. Measuring the system’s own behaviour

~5.8 new verdicts/day, backlog ~210, ~91% same-day supersession. Almost none of the 300 publish their own operating metrics — they describe what they do, not how well they do it.

The 5 biggest gaps the comparison exposes

Gap 1 — No prioritisation

The sweep processes everything equally while the backlog sits at ~210. Fixes in the field: demand-driven distillation, ASReview’s active-learning ranking, Rayyan’s AI-ranked screening. Fix: rank the queue by expected value — stale high-certainty facts first, orphans last — with a stopping rule for the long tail.

Gap 2 — Binary ratification, no confidence grading

Platinum is ratified or not. GRADE shows graded certainty is far more useful for prioritising re-verification and calibrating trust. Fix: a certainty field computed from corroboration, recency and check history — inputs the Dream Cycle already produces.

Gap 3 — No stated export or portability guarantee

Your workflows live as files on disk, which is good — but there’s no stated export format, backup, or answer to “if the vector/graph store dies, what survives?” Fix: state the format (Markdown + JSON), keep bronze and memory readable without any tool, and test a restore.

Gap 4 — Synthesis smooths over disagreement

Gold guides join silvers without checking whether they agree; meta-analysis uses heterogeneity testing precisely because pooling disagreement misleads. A gold guide could bake in false consensus before the Dream Cycle ever sees it. Fix: a disagreement check in gold synthesis — flag it in the guide and queue a verdict.

Gap 5 — Throughput measured, retrieval quality not

Verdicts/day and backlog are throughput. Unknown: whether semantic search finds the right capture, classifier precision/recall, and transcription accuracy. Fix: a small evaluation set (e.g. 20 known questions with known answers) measured monthly, classifier-correction logging, and transcription spot-checks.

The honest count

300 entries ≈ 284 independent systems

16 entries are the same underlying system under a different name, in a different track, or as a different form factor — counted openly, not hidden:

  • Zettelkasten ×4 — Track 1 maintenance; Track 2 #71 and #100; Track 3 #6. One method.
  • Tiago Forte / Building a Second Brain family ×4 — Track 2 #69 Progressive Summarization, #70 CODE; Track 3 #2 PARA, #4 BASB/CODE. Four distinct methods, one body of work.
  • Readwise ×3 — Track 2 #98 pipeline, #99 YouTube ingestion; Track 3 #38 Reader. One product family.
  • GTD ×2, Obsidian ×2, Logseq ×2, Evergreen notes ×2 — each appearing in two tracks under different comparison angles.
  • Plaud ×2 (Track 3 #14, #15 — same company and backend, different hardware), Databricks medallion ×2 (Track 2 #1, #21), MyLifeBits / SenseCam ×2 (Track 3 #9, #10 — SenseCam was a MyLifeBits component).
  • Limitless / Rewind (Track 3 #13) is one entry under two names — no count reduction, noted for clarity.

Deliberately included as verified historical systems, for the vendor-mortality lesson: Pocket (shut down 8 Jul 2025), Omnivore (hosted service shut down; self-host only), Moves (shut down 2018), Narrative Clip and Autographer (companies wound down), MyLifeBits (research project, never a product).

Not counted as duplicates: Mem0 / Letta / Zep (distinct architectures); Delta Lake / Iceberg / Hudi (distinct formats); Lambda / Kappa; Kimball / Inmon; the six meeting transcribers; the seven self-hosted photo managers; Zotero / Mendeley / EndNote; dbt / Fivetran (merged June 2026 per the sources, but distinct products).

How to read this — verification & open questions

Verification is index-level, not live-page

All 300 entries were verified from web-search results — published sources read 21–22 Nov 2026. No entry was confirmed by reading its primary documentation page in full, and no live browser checks were made. For a comparison against published systems that was judged sufficient — but any entry you plan to act on (for example, adopting a tool) should be re-verified against its primary documentation first. Each entry in the explorer links to the source it was verified from.

13 “thin” entries — existence verified, detail mention-level; unstated fields are marked “n/s” in the research notes rather than guessed: Track 1 #39 AGM belief revision, #40 Dung-style argumentation, #41 Multi-agent TMS, #48 Wiki gardening, #58 RemNote, #61 Roam Research, #62 TiddlyWiki, #63 Foam, #64 Notion second-brain review, #70 WikiProject Unreferenced Articles; Track 2 #35 Kappa architecture, #73 Commonplace book, #99 YouTube transcript ingestion.

Fast-moving facts — re-check before acting: as reported in the sources read 21–22 Nov 2026, Limitless was acquired by Meta (5 Dec 2025) and Pendant sales stopped; Pocket shut down 8 Jul 2025; Fivetran merged with dbt Labs (June 2026); Airflow 2 reaches end-of-life April 2026. These were current in those sources and may already have moved.

50 candidates were dropped rather than included without sources — 13 in Track 2 (including Lakeflow Pipelines, Informatica PowerCenter, SSIS, AWS Glue, Flink, Data mesh, Cognee ECL) and 37 in Track 3 (including tldv, Read AI, Avoma, Bear, Moleskine Pen+, Google Fit, WakaTime, Todoist, Roam capture). The count is 300 verified, not 300 at any cost.