An independent engineering case study

Reading is the start.
Returning is
the practice.

A self-hosted learning pipeline connecting readers, a durable knowledge archive, review history and personal reflections. The infrastructure keeps ideas available across devices, with a wider design for connecting learning to the projects it can inform.

Public demo: synthetic reading only. No account. No private collection.

Keep the source intact.Make selection explainable.Give reflection its own place.

Learning in context

Useful ideas.
Connected to real work.

Weave sits within a wider learning system built around existing reading tools, knowledge records, databases and workflows. The goal is to help ideas return both for deliberate review and when they can deepen understanding of a project already under way.

The current review engine is implemented. Automated recommendations based on current projects are a proposed next layer.

01 / Capture and preserve

A durable learning record.

Reader integrations and imports feed a canonical archive. Stable source identities preserve provenance while a rebuildable index keeps the review experience responsive.

02 / Review and connect

More than a list of highlights.

Source-balanced selection, separate review history, personal reflections and device synchronization form a usable learning loop. Scoped permissions, backups and readback make it maintainable.

03 / Project-aware direction

Learning for the work in front of you.

Project-linked records and earlier learning workflows provide a foundation for bringing relevant ideas to work in progress. The next stage is permissioned retrieval from project databases, with source attribution and clear user control.

01 / A connected reading practice

One collection.
Different ways in.

The goal is not another reader. It is a useful path from a passage on a device to a quiet review, and from that review to a note you can keep.

Kindle

Read without
the laptop.

KOReader with the Readest plugin. Native Readest Cloud handles acknowledged progress and notes; a WebDAV metadata mirror feeds the existing bridge.

Local reading and network delivery are separate. The bridge is designed to resume queued work while the reader is awake and connected.

Phone + laptop

Pick up
the same reading.

Readest on a GrapheneOS phone and in the web reader. Reader WebDAV sync carries books and reading state; the review app is a separate browser surface.

Reading and reviewing remain separate experiences, connected by stable source and highlight identities.

Obsidian

Keep the original.
Add your thinking.

Highlights remain source records in Markdown. Weave reads a derived index and keeps review actions separately. Reflections are exported to Notes, not written over source text.

Database save and Markdown export have distinct states and receipts.

The documented architecture

Follow a highlight, end to end.

Current Planned
System architecture. Reader devices connect to external Readest Cloud and a self-hosted WebDAV service. A metadata poller feeds a token-authenticated receiver and scoped writer, then CouchDB LiveSync and Obsidian. The separate Weave application uses a constrained connector and its own SQLite state. A future clipper has a dashed capture path.
Read the architecture without the diagram

The existing reader bridge polls WebDAV metadata read-only, sends it to a token-authenticated receiver, and passes accepted records to a scoped writer. CouchDB LiveSync carries the resulting Markdown to Obsidian.

Weave is a separate Node.js 24 application behind the existing Caddy HTTPS service. Its positively scoped connector can read Highlights and write Notes only. A cached index and durable review state live in its own SQLite database. It is not the reader bridge and does not replace the reading applications.

02 / A small, explainable review

Not everything.
Something worth revisiting.

The selector is a transparent set of heuristics, not a claim to know what you have learned. The defaults are six cards, a 35% recent-material share, and a seven-day cooldown.

  1. 01

    Respect the boundary first.

    Remove hidden, excluded, non-text, snoozed, reviewed-today and cooldown items before selection. A smaller session is preferable to silently breaking those rules.

  2. 02

    Choose sources before passages.

    Group eligible items by source before weighted random selection. In mixed sessions, take at most two from a source and reduce its weight by (1 + picks)² after each pick.

  3. 03

    Balance recent work with rediscovery.

    Reserve round(size × recency) places for eligible highlights actually created within 45 days. Import time does not make an old highlight new. Freshness and revisit signals are capped at 90 days.

  4. 04

    Let preferences nudge, not dictate.

    Source-frequency and category controls range from 0 to 3. Favorites get a modest 1.15× weight, not a mandatory place. Recent, rediscover and focused-source modes offer different starting points.

What counts as a review?

Previous / NextNavigation only. No review recorded.

Mark reviewedExplicit action. Records a review.

Snooze 7 daysAdvances without recording a review.

03 / Learning principles, carefully applied

Informed by research.
Honest about the gap.

Returning to an idea is useful only if we distinguish an interface choice from an established learning result. Weave does not measure recall or infer mastery.

01

Spacing

Cepeda and colleagues’ 2006 meta-analysis covered 317 experiments. Distributed practice can support retention; effective spacing depends on timing and the intended retention interval. It does not establish one universal schedule.

Read the primary research ↗

What that means here

A cooldown creates distance between encounters. The seven-day default is a product heuristic, not a validated optimum or a trial of this app.

02

Retrieval

Roediger and Karpicke’s 2006 experiments compared studying with free recall. Retrieval practice improved performance on delayed tests relative to repeated study, even though rereading could create greater confidence.

Read the primary research ↗

What that means here

Seeing a highlight again is not a recall test. Current Weave does not quiz you. Quizzes and measured retrieval remain future work.

03

Interleaving

Kornell and Bjork’s 2008 study examined spaced, interleaved examples in learning to recognize artists’ styles. Its structured learning task is narrower than a feed of unrelated passages.

Read the primary research ↗

What that means here

Source balancing makes a review more varied. Random highlights are not scientifically equivalent to structured interleaving, and variety alone does not prove better learning.

This is not the Free Spaced Repetition Scheduler (FSRS), an adaptive mastery model, or a reproduction of a proprietary review algorithm. It is inspectable weighted selection.

Selection + explicit actions

A review is an action, not a page view.

Review flow from immutable highlight records through a derived eligible pool, source grouping, weighted selection and the daily browser session. Only Mark reviewed writes review history. Reflections save to SQLite before a separately acknowledged Notes export.

04 / Small permissions. Visible boundaries.

The demo is public.
The collection is not.

Two environments share an interface and selection logic, not a data connection. The public version demonstrates the product without exposing the private system.

This public demo

A browser-local
sandbox.

  • Exactly ten original, general-topic passages from five fictional books or articles.
  • The real interface and selector, with a browser-local data adapter.
  • No backend, login, cloud sync, artificial intelligence (AI) calls or tracking scripts.
  • Actions stay in this browser’s local storage. Reset affects only this visitor.
  • No reader bridge, private connector, or personal collection access.

The static host may retain ordinary request logs. Demo notes and actions are never uploaded.

Try it for yourself ↗

The private application

Scoped access.
Separate state.

  • Node.js 24, native SQLite and a vanilla browser interface.
  • A separate, positively scoped connector: read Highlights; write Notes only.
  • Non-root containers, read-only container filesystems, scoped tokens and existing Caddy HTTPS.
  • Daily consistent SQLite backups, fourteen retained; isolated restore tested.
  • The host is trusted. This is not end-to-end encryption against its operator.

A documented design, not a fresh production audit. The first scheduled backup execution has not yet been observed.

Separate trust boundaries. A visitor browser loads static demo assets and uses only its local storage. The private application has its own HTTPS service, SQLite storage and restricted connector. There is no connection from public demo to private data. The trusted self-hosted operator can read private data.

“Saved” should say where.

A reflection first receives a durable database save. Markdown export to Notes is a separate step with a pending state, delivery receipt and read-back. Saving locally must not be presented as confirmed delivery to Obsidian.

Planned concept / not an extension

The next way in:
a deliberate clip.

A web passage should be able to join the same collection without a second inbox to maintain. The intended flow starts with a selection and an explicit Save, not background page collection.

  1. Select a passage on a page.
  2. Preview the destination, title, source URL and selected text.
  3. Save once with provenance and an idempotent capture identifier.
  4. Keep capture reversible, without duplicating repeated saves.

Minimal permission on activation, rather than access to every site. Private, unpacked distribution is planned. The extension and capture service are not implemented.

an article you chose
FIELD NOTES / DESIGN

Notice what
you keep.

A clear explanation preserves the decision, the alternatives, and the reason one of them was chosen.

PLANNED CONCEPT

Keep this passage

Destination
Highlights collection
Included
Selected text + source + title
Preview → Explicit Save

Illustration only. Nothing is captured.

05 / Choices behind the interface

Small on purpose.
Not simple by accident.

A derived index, not a new source of truth.
A cache makes review responsive. Immutable source records preserve provenance; review history and reflections have their own durable state. The cost is explicit synchronization and export states.
A narrow connector, not an open vault.
Read Highlights and write Notes only. This reduces the application’s reach, but does not remove host trust or the need to check the receiving system.
Heuristics, not hidden intelligence.
Deterministic tests can exercise the selector’s boundaries. An explainable default is easier to inspect, at the cost of not adapting to measured memory or predicting mastery.
A dedicated review app, not another reader.
Readest and KOReader remain the reading surfaces. Weave focuses on highlights now. Full offline Weave, Kindle review, AI discussion and project-learning agents are deferred.

06 / Evidence, not implication

What exists.
What still needs proving.

A working configuration, a successful test and acceptance on a real device are different states. This snapshot keeps them separate.

Documented project state accompanying release 0.2.3. This page does not claim a new live-server verification.
Capability State Evidence or remaining work
Private web review app Implemented Production release 0.2.3. Recorded suite: 65 tests; five responsive viewports passed.
Phone / real-device acceptance Follow-up open Responsive checks are not acceptance on the actual phone. Device follow-up remains open.
Kindle reading bridge Integration tested Recorded reading, highlight and retry checks passed. Ongoing device health and fresh-delivery acceptance are maintained separately.
SQLite recovery Restore tested Consistent backup + fourteen-copy retention configured; isolated restore tested. First scheduled execution not yet observed.
Reflection to Notes Explicit states Database save first. Export can remain pending until delivery receipt and read-back confirm it.
Public demonstration Static + local Synthetic fixtures and local browser state only. No private-system connection.
Web clipper Planned concept Preview and explicit-save design only. No extension or live capture implementation.
Kindle review, offline Weave, quizzes, AI discussion Deferred Not deployed features. The current focus is highlight review.

Built to be inspected, not taken on faith

A quieter interface.
A clearer system underneath.

Use the demo to explore the experience. Use the diagrams and explicit limits to judge the engineering.

Spend a moment with Weave