Dinesh/ Blog
learning

Building a Learning System That Actually Sticks

Research-informed retrieval and spacing with a broad curriculum, honest schedule arithmetic, bounded review queues and no invented automation or mastery promises.

30 min read
Checking device speech…

In this article

  • Distinguish recognition from unaided retrieval
  • Build a failure-driven review queue with explicit retests
  • Separate a scheduling heuristic from evidence of mastery
Background & source details

Level: beginner

Learning and deliberate practice · Lesson 1 of 1 ↗
Loading this browser’s progress…

Sources checked: . A source check does not change the publication or update date.

Editorial review: · What review means

In this article · 27 sections

Evidence and scope

This is a proposed study system, not an account of a completed personal experiment. Its retrieval and spacing choices are informed by research; its calendar, time budgets and queue labels are adjustable engineering conventions. No calendar event, reminder, nightly job or progress record is created by reading this page. The source inventory contains 130 numbered engineering sessions, 65 numbered algorithm-practice sessions and 26 weekly revisions—not 27 revisions. Those counts describe source documents, not completed study or independently accepted runtime demonstrations.

Building a Learning System That Actually Sticks

A familiar failure is to accumulate notes without being able to explain or use their contents later. That can reflect shallow encoding, missing prerequisites, insufficient retrieval, unsuitable practice, overload or a poor schedule. It is not evidence that the learner lacks discipline, nor proof that more effort could never help.

The design below makes learning observable: attempt something, inspect the error, repair it, then test again after a delay on a changed example. It preserves a broad curriculum rather than replacing difficult subjects with the few tasks that fit a convenient hour. A manageable schedule is useful precisely because it can expand when an objective needs more work.

Part 1 · Why studying feels productive and isn't

The fluency illusion

Familiar wording can be easy to process without being easy to reconstruct. Rereading is not useless: it can supply an explanation you never understood or correct an error. But the feeling “I recognize this” answers a different question from “Can I derive and apply this without the page?” Test the latter directly.

In Roediger and Karpicke’s 2006 experiments, repeated study helped on the very short-delay test, while prior testing improved delayed retention. This is a performance-versus-retention distinction, not a rule that unpleasant sessions are better. Desirable difficulty is conditional: a challenge must be achievable with suitable feedback and prerequisites. Confusion, sleep deprivation and repeated uncorrected errors are not evidence of learning.

Practical consequence: after a smooth explanation, try a closed-book example. After a frustrating attempt, inspect whether you lack a prerequisite or need a smaller example. Judge both by later performance, not by how the session felt.

What the evidence actually supports

  1. Retrieval practice. The 2006 passage-learning experiments found a delayed benefit from taking initial tests relative to restudy. They do not establish that testing always beats explanation, that feedback is unnecessary, or that the result directly predicts coding interviews. Combine retrieval with corrective feedback and an application task.
  2. Spacing. Cepeda and colleagues’ 2006 meta-analysis concerns distributed practice in verbal recall. Study gaps and final retention intervals interact. It supports separating practice in time, not a universal exponentially expanding schedule or the constants +2/+7/+21/+60. The target delay, material and learner matter.
  3. Interleaving. Rohrer and Taylor’s 2007 mathematics experiments provide evidence for shuffled practice under their conditions. Mixing distinguishable problem types can require selecting a method rather than merely repeating it. The authors explicitly limit generalization: their college-student laboratory tests used problems like those practiced, not a transfer measure, and procedural rather than abstract conceptual tasks. This does not prove that arbitrary topic-switching is beneficial or that initial blocked instruction is wrong. Learn a new technique with support, then contrast it with alternatives.
  4. Attempt before answer. Kornell and colleagues’ 2009 experiments studied unsuccessful retrieval followed by learning the answer. The feedback/answer stage matters. An unsupported guess is not automatically a useful memory trace; a novice may first need definitions and a worked example.

No percentage improvement, exact session length, guaranteed mastery date or motor-skill classification follows from these studies. A useful local check is to compare two study methods on similar material using a delayed unseen task, while recognizing that one learner’s small comparison is not a controlled research replication.

The engineering translation

Research-informed ideaProposed implementationCheck and failure response
RetrievalEnd with a closed-book explanation or reconstructionCompare with a reference; correct omissions before repeating
SpacingAssign a later retest instead of doing every repetition todayShorten an excessive gap; extend one that yields only immediate familiarity
DiscriminationMix previously introduced problem typesRequire an argument for the choice, not just a pattern name
Generation with feedbackPredict before revealing the answerIf prerequisites are absent, teach them rather than prolonging guessing

These are design choices derived from findings, not experimental validation of this whole workflow.

Part 2 · The architecture

Use three complementary tracks plus time for consolidation:

  • Concepts: 130 sessions across 16 modules, from setup and Python through mathematics, data, systems, ML and AI design. The full curriculum retains every numbered destination.
  • Algorithm practice: 65 lessons in five phases, from contracts through DP and advanced structures to mixed assessment. The family hub is broader than the ten-lesson Foundations Lab.
  • Revision: 26 weekly articles connecting groups of five concept sessions. Revisions include worked exercises and explicit execution limits; their number is not a completion record.
  • Consolidation: leave slack for rest, a corrected note and a later retrieval attempt. Summarizing notes is information organization, not evidence that their author learned them.

Why separate tracks at all

Explaining a B-tree and solving a new algorithm problem both involve conceptual knowledge, retrieval, reasoning and practice. Typing is a motor activity, but algorithmic problem solving is not merely motor memory. There is no established decay-rate dichotomy here that licenses a different calendar for “semantic” versus “motor” topics.

Separate tracks are an organizational choice: a concept session can develop a mechanism and its tradeoffs; a practice session can emphasize choosing and implementing it under changed constraints. Combine them if that reduces overhead. Keep their evidence distinct: a narrated explanation, a passing program, a proof and a timed solution are different achievements.

The schedule

One internally consistent optional 26-week calendar is:

ActivityFrequencyDurationWork over 26 weeks
ConceptsFive each week60 minutes130 hours
AlgorithmsAlternate 3-session and 2-session weeks75 minutes81.25 hours
RevisionOne each week90 minutes39 hours

The total is 250.25 hours, averaging 9.625 hours/week (9 hours 37 minutes 30 seconds), before breaks, note maintenance or overruns. Three-algorithm weeks take 10.25 hours; two-algorithm weeks take 9 hours. This is not “every other calendar day”: that would produce about 91 algorithm sessions in 182 days, not 65. For a concrete weekday policy, use Mon/Wed/Fri mornings in odd weeks and Tue/Thu mornings in even weeks. Five evening concept slots and a Saturday revision are examples, not claims about an existing calendar.

Morning practice may avoid workday overruns but can reduce sleep or clash with caregiving. Evening practice may fit better but compete with fatigue and other tasks. Choose from actual availability and attention, not the unsupported premise that everyone has peak executive function after waking. A missed task enters a bounded make-good queue; it does not automatically consume the already full revision block. Replace a lower-priority task or extend the end date. Preserve rest instead of doubling tomorrow’s work indefinitely.

Interleaving, concretely

The 16 modules include setup, Python, mathematics, DSA, databases, data engineering, backend, infrastructure, distributed systems, observability, security, ML, DL, NLP/transformers, LLM applications and system design. Keep this breadth; vary the sequence only after checking prerequisites.

A small illustrative braid is environment → variables → Git → control flow → command line → HTTP → functions → REST. It mixes tooling, Python and backend, but it is not a published 130-node schedule or a validated seven-track calendar. For a learner who cannot yet read request examples, HTTP may belong later. Interleaving previously learned search and window problems directly exercises method selection; alternating unrelated introductory subjects does not necessarily do so.

The site’s read-only graph distinguishes editorial sequence (plan/order) from explicit prerequisiteSlugs. Its validator checks duplicate positions, missing public prerequisite targets and cycles. It does not generate calendar events, infer all conceptual prerequisites or prove that a round-robin order is suitable. Before manually reordering lessons, list required skills, check the named dependencies, and use a diagnostic to catch missing textual prerequisites. A prerequisite DAG constrains allowed orders; it does not choose study durations.

Part 3 · The session protocol

The protocol is a template you may put in a calendar event. It is not already installed there. Use one coherent example from intuition through code and explanation; postpone advanced branches rather than pretending to have finished them.

The 60-minute concept protocol

StageMinutesArtifact
Blank page2A prediction and one uncertainty; definitions first if needed
Intuition8Explain what problem the mechanism solves
Visual10Redraw a state, memory or request trace and label transitions
Code or derivation20Run a safe local example or work a source-supported calculation
Production reality8One documented failure mode, assumption and tradeoff—not an invented war story
Retrieval close12Explain, compare with reference, and record one retest

The blocks sum to 60 minutes. A cloud deployment, difficult proof or full debugging session may need separate time and authorization. Do not run unsafe commands to satisfy the timer.

Close with three questions:

  1. What did I learn? Give the mechanism without copying its words.
  2. What breaks without it? Identify the requirement it serves and a simpler alternative.
  3. Where could it apply? Use a genuine experience if you have one; otherwise label a hypothetical example.

For example, write-ahead logging can be connected to recovery after an interrupted update. That is a technical scenario, not a claim that the author debugged a particular pipeline or Copilot system. State the log-before-data rule and ask what the recovery procedure can reconstruct.

Encoding versus retrieval mental model: a warehouse analogy distinguishes storing a box from finding it, but it is not a literal model of memory. Poor initial understanding can matter as much as missing retrieval cues. Rereading can repair that understanding; retrieval exposes what is accessible; useful prior knowledge supplies connections; spacing changes the conditions of the next attempt. When about to reread, first ask one question you want the rereading to resolve, then close the page and answer it.

The 90-minute Saturday revision protocol

Choose a bounded subset of new and older objectives, rather than promising to reconstruct everything studied so far each Saturday. One alternative to the weekly article’s own 90-minute agenda is: 20 minutes closed-book recall, 15 checking gaps, 30 rebuilding one example, 15 writing question/answer cards, 10 connecting it to older material. Use one agenda, not both; each totals 90 minutes.

Keep correct answers as well as gaps so review does not become a list of permanently familiar failures. A blank page that stays blank is a reason to adjust the task and supply feedback, not proof that discomfort is doing its job.

Misconception: cards cannot teach engineering. A definition card can support vocabulary, but cannot by itself assess an implementation or design. A reconstruction card can ask for mechanism and assumptions. Compare “B-tree lookup is logarithmic” with “Why can a page-oriented B-tree need fewer page reads than an unbalanced or pointer-heavy binary tree?” An answer should mention high fanout, page-sized nodes and reduced height, then qualify caching and workload. Neither card replaces an index experiment. Give the answer and rationale on the reverse side; questions without feedback can preserve mistakes.

Part 4 · The algorithm track

The 65 lessons retain the full problem-family route. The 75-minute format is a pacing option, not the amount of practice that guarantees an outcome.

Pattern-first, not problem-first

Initially grouping examples can make an invariant visible. Later mix families so the lesson title no longer supplies the method. Do not infer a universal advantage for “150 structured problems” over “500 random problems”; no such experiment is established here. Progress is better assessed by transfer, correctness and explanation than by a solved counter alone.

Constraints are hints, not a decoder for the setter’s intended algorithm. At n = 100_000, an all-pairs loop examines about five billion unordered pairs; an n log2(n) count is about 1.66 million. Constants, Python allocation, input representation and the actual operation still matter. A claim of 100 million Python operations per second is not a portable budget. Small n can make subset enumeration feasible but does not make n! feasible: 20! is about 2.43 quintillion. A numeric value of a billion may allow a logarithmic search over values, whereas a billion explicitly supplied inputs already impose an input-reading cost. Count the actual work and test correctness separately from runtime.

Structure

  • L01–L06, foundations: contracts, operation counts, Python containers, mutation discipline, hashing and UMPIRE. Output: a baseline plus explicit assumptions.
  • L07–L32, core patterns: pointers, windows, prefixes, binary search variants, stacks/deques, lists, recursion, trees, heaps, graph traversal/ordering/connectivity and backtracking. Output: the invariant and a counterexample if its precondition is removed.
  • L33–L44, dynamic programming: recurrence → memoization → table; linear, take/skip, subsequence, knapsack, grid, string, interval, state-machine and bitmask families. Output: state sufficiency, transition order and a reconstruction or oracle check.
  • L45–L56, advanced subjects: tries, bit and number reasoning, intervals, greedy exchange arguments, matrices, LRU/stream structures, comparators, range trees, string matching and weighted graphs. Output: a defensible choice and its limits.
  • L57–L65, assessment: mixed sets, narration, gap audit, targeted repair, hard decomposition, company-tagged transfer, senior-style composition, contest and maintenance. Output: a truthful attempt log, not a hiring guarantee.

The two-attempt rule

Use one consistent optional policy across this essay and the hub: try unaided for up to 15 minutes, request a small hint if useful, then spend up to 10 more minutes deriving the solution. If still stuck, read the worked solution, close it and reconstruct with feedback. Schedule a fresh attempt about two days later or at your next feasible review slot. These are stopping rules to control unproductive struggle, not research-derived constants. Record whether a pattern-name hint supplied the very recognition being assessed.

For a 75-minute session, budget 5 warm-up, 10 anatomy, 15 guided solve, 25 solo work, 10 post-mortem and 10 spaced review. That is 75, not the original 80-minute sum. Three full problems may not fit; problem variants can be discussed without falsely counting them as independent solves.

Status codes drive the spacing

StatusMeaningIllustrative next review
coldUnaided solution with a valid argument and testsUp to +60 days, earlier if transfer is untested
warmSolved, but slow or with a repaired error+21 days
hintNeeded external guidance+7 days
failedNeeded a worked answer+2 days after correction

These intervals are optional queue defaults. They do not establish mastery, and a successful familiar problem should not automatically defer all variants for 60 days. Store an actual due date based on elapsed days, not a session-number offset: L09 minus L05 is four sessions, not necessarily seven days. No existing event is asserted to carry this queue. On overload, prioritize an important recent gap and one older transfer problem, then explicitly reschedule the rest.

The failure log

Record the missed cue and the failed contract, state, implementation or cost argument. A hypothetical entry: “I chose earliest start for maximum-cardinality non-overlapping intervals; [0,10), [1,2), [2,3) refutes it. For the unweighted problem, choosing earliest finish admits an exchange argument; weighted intervals need a different rule.” This is more useful than “weak at greedy,” and does not assume every failure is recognition rather than coding.

A later retest should change the intervals or introduce weights. Explain which previous proof step stops working. Keep the failed input, corrected reasoning, exact code/test and whether the later attempt was unaided.

Part 5 · The content layer

A curriculum needs more than a route list. Four editorial checks help make its explanations inspectable; they are not claims that an audit found zero examples across 210 files or that four new UI components were built:

  • Misconception: state the tempting wrong rule, why it seems plausible, a counterexample and the corrected boundary. For aliasing, contrast immutable integers with nested mutable lists; do not simply tell the learner “copying is broken.”
  • First principles: expose the assumptions and derive the mechanism. A derivation offers a reconstruction path, but understanding and later practice are still required.
  • Tradeoff: name the cost and a testable decision condition. “When appropriate” is weak; so is an unexplained universal 10k-write/s threshold. Prefer “when the measured single-node workload violates this latency target after the stated tuning, and this partition key avoids the measured hot spot.”
  • Mental model: use a compressed picture and the situation that should trigger it, then say where the analogy fails. A model is not guaranteed to survive six months without use.

The current reading surface may express these as plain sections, tables or exercises. Correctness does not depend on a particular MDX component name. Each substantial objective needs an explanation and a worked or reasoned answer, not a heading with a generic instruction to practice.

Part 6 · Consolidation and the memory layer

Sunday is load-bearing

Sunday is slack in the sample calendar, not a biologically privileged day. Sleep-and-memory research supports an important role for sleep in consolidation, while effects vary across tasks and memory systems. It does not prove that studying on seven days prevents consolidation or that studying near bedtime is harmful. Protect sufficient sleep and recovery daily. Use Sunday for rest, a short enjoyable application or a moved task only if that fits your needs; do not silently add a seventh full workload.

Automated nightly consolidation

A possible note workflow has a raw attempt log, a dated digest and a small curated index. The proposal to summarize first and retire working context later is an information-preservation rule, not a biological sleep analogue or a claim that 02:00/02:15 jobs exist.

For safe implementation, write a digest to a temporary file, validate its references to the original log, atomically publish it, and only then update the index. A failed digest must leave the source log available. Use stable IDs so rerunning a job does not duplicate entries. Keep raw attempts separate from polished model summaries: a summary can introduce errors and cannot mark a skill learned.

Compression tradeoff: aggressive summaries reduce reading cost but omit detail; a hot index plus a searchable cold archive preserves provenance at the cost of storage, indexing and privacy management. “Never delete” is not an unconditional rule: sensitive data, consent and retention requirements can require deletion, and backups can fail. Set explicit retention and access policies, minimize private work content, test restoration, and cap the hot tier rather than pretending storage is free. Archiving originals makes compression inspectable, not infallible.

Part 7 · What I would tell someone copying this

Start with both a small diagnostic and a realistic time budget; neither an excellent curriculum without practice nor an easy schedule with incorrect content is enough. Put a short protocol in the event if you use a calendar. Make the next review easy to find, but let workload and observed gaps change it. Track specific failures, not a global verdict about yourself.

Expect some effort, not a required month of misery. Repeated inability to start is a signal to revisit prerequisites, reduce task size or obtain instruction. A streak can motivate some learners; a make-good queue can help others. Neither should reward fabricated completion. Mark “read,” “attempted,” “corrected” and “retested unaided” separately.

The honest limitations

This article reports neither four personal curriculum restarts nor a completed August-to-December trial. It does not claim that 250.25 scheduled hours suffice for the whole breadth, that each component has been tested together experimentally, or that all linked lessons have completed release acceptance. The source-backed research is narrower than the proposed workflow.

If work pressure collapses the schedule, reduce concurrent tracks or extend the plan. A bounded queue with priorities and explicit rescheduling is more honest than an ever-growing hidden backlog. The success criterion is retained, transferable understanding demonstrated on suitable tasks, not finishing by a date.

Appendix · The stack

LayerPossible implementationBoundary
CurriculumMarkdown under version controlPreserve source and review changes; git is not access control for secrets
LessonsMDX reading surfaceRendering is not learning verification
ScheduleLocal calendar or an optional API integrationGenerating a plan is pure computation; applying events is an external side effect
Session dataJSON with stable IDs and prerequisitesValidate types, duplicates and missing dependencies
Review queueStored due dates plus a bounded selection ruleElapsed days and lesson offsets are different
ProgressA table of attempts and retestsRecord evidence, not page visits as mastery
ConsolidationManual routine or separately configured automationIdempotence, failure recovery, privacy and backups require tests

No cloud account or paid tool is required. A proposed calendar generator should support dry-run diffs, stable event IDs, time zones, conflict detection and idempotent updates; “regenerate in one command” must not mean blindly duplicating or deleting events. Keep private notes out of public repositories. Copy the workflow as a hypothesis to adapt, not a personal history or service that this page claims to have installed.

Corrective source notes

Research references above identify the actual experiments or review and their limits. The site’s current learning-graph.ts is the source for the graph description; a graph validation rule is not proof of a deployed scheduler. Counts and schedule arithmetic were checked against the source inventory on 2026-10-08. This review covers the essay and its curriculum-navigation claims, not a fresh substantive review of every destination.

Pause / Recall / Apply

Can you explain it without the page?

Close the example. Reconstruct the core idea, then change one assumption. Mark complete when you’re ready; you can always undo it.

Stored in this browser only. No account, no sync. Clearing browser data removes your record.