Mnemosyne — Features

The humanistic learning and cultural intelligence platform

Mnemosyne is named after the Greek Titaness of memory — mother of the nine Muses and the source of all artistic inspiration. The platform provides an integrated system for language acquisition, memory science, classical education, art history, mythology, cultural heritage preservation, history, and archaeology. It unifies scientifically grounded spaced repetition and adaptive learning algorithms with deep coverage of the humanities: from learning Ancient Greek vocabulary with morphological analysis support, to studying iconography in Renaissance painting, to preserving endangered cultural practices through digital archiving.

The domain ships as 19 pure TypeScript libraries. Each package is a self-contained computation library: there is no database, no HTTP server, and no network client inside libs/mnemosyne/*. Every capability described in this document is implemented as deterministic functions and in-memory data structures that a consuming application can call directly.

Domain Libraries#

The table below gives a one-line description of each library's role. Each is described in depth in its own section below.

Memory Science and Spaced Repetition (@mnemosyne/core)#

Spaced repetition is the scientifically validated method of reviewing information at progressively longer intervals to maximize long-term retention. Rather than committing to a single algorithm, Mnemosyne implements multiple memory models, each making different assumptions about how human memory works. The learner or the application can choose the model best suited to the content and context.

Forgetting Curve and Retention Models#

Ebbinghaus forgetting curve is the foundational memory model, representing retention as exponential decay from initial learning. Mnemosyne uses the formula R = e^(-t/S) where S is the "stability" of the memory and t is elapsed time, with per-card half-life estimation based on review history.

SM-2 algorithm is the original SuperMemo algorithm (1987) that powers Anki and most spaced repetition software. It uses an "ease factor" (starting at 2.5) adjusted up or down based on answer quality (0–5 scale), with the next interval calculated as interval × ease_factor. The SM-2 grade scale distinguishes a complete blackout (0) from a perfect response (5); answers below 3 trigger relearning.

FSRS (Free Spaced Repetition Scheduler) is a modern 17-parameter model that explicitly tracks both memory "stability" (how long it takes to forget) and "difficulty" (intrinsic item hardness) as separate dimensions. Unlike SM-2, FSRS predicts the probability that a learner will recall an item at any future point, enabling configurable target retention (default 90%). FSRS outperforms SM-2 especially for irregular review histories and forgotten items.

Half-Life Regression (HLR) is Duolingo's research model that fits a personalized forgetting curve per learner per item using logistic regression on review history. It tracks hlrPredict (retrieval probability at future time) and hlrRetention (current estimated retention).

LECTOR scheduling is a semantic-aware scheduling algorithm that accounts for conceptual similarity between items — avoiding scheduling similar vocabulary on the same day to reduce interference effects.

Optimal Review Scheduling#

Beyond individual item scheduling, the system provides session-level planning tools:

• Next-review time calculation: Balances predicted forgetting probability against total daily review load to prevent overwhelming review queues.
• Review prioritization: Ranks items by combined forgetting probability and learning value; items near the forgetting threshold rank higher than items nearly forgotten.
• Daily review load forecasting: Projects review counts N days ahead for planning purposes.
• Interleaving strategy generation: Schedules items from multiple knowledge domains in the same session to exploit the interleaving effect — which produces better long-term retention than blocked practice despite feeling harder.

Cognitive Load Management#

Cognitive overload is a real barrier to effective study. These features prevent it:

• Session cognitive load estimation: Estimates total cognitive demand based on item difficulty distribution and recent study history, preventing sessions that are counterproductively exhausting.
• Session length recommendations: Caps that prevent overload while maximizing retention per hour.
• Circadian rhythm-aware scheduling: Uses getCircadianPhase and circadianEfficiency to recommend study at times of day when cognitive performance is highest for the learner's chronotype.
• Sleep-aware scheduling: sleepAwareSchedule avoids scheduling difficult items when the learner is likely fatigued; hoursUntilSleep determines whether a session is in the optimal pre-sleep consolidation window.

SRS Card System#

A single SRSCard carries both SM-2 state and FSRS state so the two algorithms can operate on the same card object without duplication:

• Dual-algorithm card model: One card holds SM-2 fields (easeFactor, interval, repetitions) and FSRS fields (difficulty, stability, state, reps, lapses). The functions sm2Review and fsrsReview both accept and return an updated SRSCard.
• FSRS state machine: fsrsReview drives the New → Learning → Review → Relearning transition graph on the card's state field.
• Deck reference: Each card carries a branded DeckId. Hierarchical subdeck nesting exists in @mnemosyne/platform's AnkiDeck interchange type, not in the core SRSCard model.
• Review forecast projections: calculateReviewLoad projects daily review counts N days ahead (ReviewForecast).

Assessment Framework (@mnemosyne/core)#

Item Response Theory (IRT)#

Item Response Theory models the probability of a correct answer as a function of both item difficulty and learner ability. This provides more precise measurement than raw score percentages because it separates what the test measures from how hard the items happen to be. Three increasingly realistic models are implemented:

• 1PL Rasch model (irt1PL): The simplest IRT model, parameterizing items by difficulty only. Used when item discrimination is assumed uniform — appropriate for carefully constructed item banks.

• 2PL model (irt2PL): Adds discrimination to the 1PL model — a high-discrimination item sharply distinguishes between learners near the threshold; a low-discrimination item does not. More realistic for heterogeneous item banks.

• 3PL model (irt3PL): Adds a pseudo-guessing parameter — the probability of a correct answer even for a learner with negligible ability. This is important for multiple-choice items and is the most realistic model for typical test conditions.

• Item information functions: itemInformation measures how precisely a single item estimates ability at a given theta level; testInformation aggregates precision over a full test. Both are essential for test design optimization.

• Ability estimation: estimateAbility via Maximum Likelihood Estimation (MLE) and Expected A Posteriori (EAP) Bayesian estimation. standardErrorFromInformation translates the test information function into confidence intervals on ability estimates.

Computerized Adaptive Testing (CAT)#

Adaptive testing selects each successive item based on what has been learned from all previous answers — narrowing the estimate of the learner's true ability as efficiently as possible rather than administering a fixed set.

• Maximum information item selection: Always selects the item that would reduce uncertainty about the learner's ability most, given the current estimate.
• A-stratified and progressive-restricted selection: Alternative selection strategies that balance information gain with item exposure control.
• Configurable stopping rules: Standard error threshold (terminate when precision is sufficient), minimum/maximum item count (hard session length limits), and minimum information (avoid items contributing negligible measurement value).
• Content balancing: Ensures the adaptive test covers required content domains in specified proportions.
• selectNextItem, shouldTerminate, runAdaptiveTest: A complete CAT loop from initial item selection through termination.

Rubric Assessment#

Rubric-based assessment handles open-response submissions where a simple right/wrong grade is insufficient:

• Rubric criterion and level definition: Multi-criterion rubrics with explicit performance descriptors per level.
• evaluateWithRubric: Scores open-response submissions against defined criteria.
• generateFeedback: Produces criterion-level feedback explaining each score.
• Portfolio assessment: Evidence submission, reviewer assignment, and portfolio evaluation workflow.
• Peer review system: Peer assignment, submission, and calibration scoring to ensure inter-rater reliability.
• Self-assessment calibration: Tracks accuracy of learners' self-predictions and improves metacognitive accuracy over time.
• Certification management: Requirement definition and credential issuance when all requirements are met.
• Proctoring integrity scoring: computeIntegrityScore for assessment integrity monitoring.

Knowledge Graph Infrastructure (@mnemosyne/core, @mnemosyne/knowledge-graph)#

The knowledge graph connects concepts across the humanities domains, enabling prerequisite-aware learning paths and cross-domain discovery. Rather than treating each fact in isolation, the graph makes explicit that knowing Latin grammar is a prerequisite for reading Caesar, or that understanding Byzantine iconography connects to studying Orthodox theology.

• KnowledgeGraph class: A directed graph for knowledge items and their relationships, with typed nodes and edges.
• Relationship types: Prerequisite (must know A before B), related (A and B are connected), contradicts (A and B are in tension), part-of (A is a component of B), exemplifies (A is an example of B).
• calculateSimilarity: Semantic similarity between knowledge items using embedding distance.
• SPARQL-inspired query interface: TriplePattern and GraphQuery for expressive semantic queries over the graph.
• Learning path inference: getPrerequisites, getDependents, and topologicalSort compute study sequences from prerequisite-graph traversal. identifyKnowledgeGaps finds missing prerequisites between a learner's current state and a target competency.
• Cross-domain concept linking: discoverCrossDomainLinks creates semantic_similar edges between similar nodes in different domains — connecting concepts across humanities domains automatically.
• Graph-based recommendation engine: recommendNextItems surfaces items whose prerequisites are met, ranked by dependent count and mastery gap. findSimilarNodes surfaces related concepts by property similarity.
• Temporal knowledge graph support: getTimeline, getNodesInTimeRange, and buildTemporalChain traverse and build temporal edges, ordering nodes chronologically by a year/date property.

AI Tutoring Infrastructure (@mnemosyne/core)#

The AI tutoring layer is designed to work with or without a live LLM. With an injected LLMProvider it can use generative power; without one it falls back to template-based and algorithmic implementations. This means the domain libraries remain useful even in environments where calling an LLM is not available.

Socratic Dialogue System#

The Socratic method guides learners toward discovery rather than simply providing answers. This is implemented across several cooperating functions:

• Socratic question generation: generateSocraticQuestion creates targeted questions designed to guide learners to discover the answer themselves. createSocraticDialogue and advanceSocraticDialogue manage multi-turn Socratic exchanges.

• Four-level hint sequences: generateHint and createHintSequence produce graduated hints across the HINT_LEVELS scale — nudge, clue, explanation, solution — from the most oblique nudge to the near-direct solution. getNextHint advances the sequence and selectInitialHintLevel chooses the starting level from learner mastery and problem difficulty.

• Multi-perspective explanations: generateMultiPerspectiveExplanation produces seven explanation perspectives for the same concept — analogy, formal, visual, historical, practical, first_principles, and comparative — then selectBestPerspective chooses the format most effective for a particular learner's cognitive style.

Misconception Detection#

Common mistakes in humanities learning often follow predictable patterns (e.g., confusing Latin cases, misattributing artworks, or applying anachronistic historical frameworks). The misconception detection system catches these proactively:

• COMMON_MISCONCEPTIONS: A curated library of domain-specific misconception patterns covering grammatical, etymological, and historical error types.
• detectMisconceptions: Flags probable misconceptions in learner responses by pattern-matching against the misconception library.
• createMisconceptionPattern: Extends the library with new domain-specific patterns as they are identified.

AI Content Generation#

Rather than relying entirely on hand-authored content, the AI layer can generate study materials from arbitrary text input:

• Question generation: generateQuestionsFromText automatically produces multiple choice, short answer, true/false, and matching questions from any text.
• Distractor generation: generateDistractors creates plausible-but-incorrect answer choices for MCQs — the quality of distractors determines whether a question actually tests understanding.
• Cloze deletion: generateClozeDeletions creates fill-in-the-blank exercises; estimateClozeDifficulty predicts difficulty based on word frequency and syntactic position.
• Semantic flashcard generation: generateSemanticCards and generateCardsFromText produce well-formed study cards from text input.
• Personalized examples: generatePersonalizedExamples generates examples of target grammar or vocabulary in contexts matching the learner's declared interests.

Learning Analytics#

Early detection of learning problems enables intervention before engagement collapses:

• Dropout risk prediction: predictDropoutRisk detects early warning signals of disengagement before the learner stops entirely.
• Mastery timeline prediction: predictMasteryTimeline forecasts when the learner will reach target proficiency given current study pace.
• Optimal review count prediction: predictOptimalReviewCount estimates the minimum number of reviews needed to reach durable mastery.
• Engagement trend analysis: analyzeEngagementTrend distinguishes between temporary dips and persistent disengagement patterns.

Multi-Language Mastery Engine (@mnemosyne/polyglot)#

@mnemosyne/polyglot is the primary home for modern-language learning features. It covers vocabulary, grammar, reading, and proficiency measurement, with particular attention to typologically diverse languages.

Language Metadata#

Before any instruction, the system needs to understand the structure of the target language:

• Comprehensive language database: ISO 639 language codes, language families and subfamilies, writing system metadata (Unicode ranges, directionality, morphological type, word order).
• Part of speech definitions per language: Language-specific POS taxonomies reflecting each language's actual grammatical categories.
• Frequency word lists: Corpus-derived word frequency lists per language — essential for teaching high-utility vocabulary first.
• Cognate databases: Cross-language related words to leverage transfer knowledge from languages a learner already knows.
• False friend databases: False cognates that mislead learners (e.g., Spanish "embarazada" = pregnant, not embarrassed).

Vocabulary Acquisition#

Vocabulary learning in Mnemosyne is not isolated drilling — it is organized around meaningful context and family relationships:

• Vocabulary item model: Word forms, definitions, and register (formal, informal, slang, technical, literary).
• Thematic module organization: Vocabulary grouped into contextual sets (travel, food, emotions, business, etc.) for situational learning.
• Word family derivation tracking: Root → all derived forms (e.g., educate → education, educational, educator, miseducate), enabling family-based vocabulary expansion.
• Mnemonic keyword generation: The keyword method — creating a memorable visual or phonetic bridge between a foreign word and its meaning.
• Vocabulary coverage analysis: Percentage of a target corpus covered by known words. The 95% threshold is the practical level for comfortable reading; this metric allows setting realistic milestones.
• Thematic activity design: Contextual vocabulary practice within thematic scenarios rather than isolated drilling.

Grammar Instruction#

Grammar is represented both as reference material and as generative exercises:

• Grammar rule definitions with examples: Structured rule representations with positive and negative examples.
• Morphological paradigm tables: Conjugation and declension tables per language for reference and drilling.
• Syntactic pattern definitions: Slot-filling representations of sentence structures.
• Grammar exercise generation: Fill-in, transformation, and translation exercises generated from rule specifications.
• Contrastive analysis: Identifies L1→L2 transfer errors based on structural differences between the learner's native language and the target — predicting which mistakes are most likely before they occur.
• Error pattern libraries: Documented common mistake types per L1–L2 pair.

Reading Comprehension#

• Text difficulty analysis: Readability metrics, vocabulary profile, and Lexile equivalent.
• Comprehension skill taxonomy: Main idea, inference, vocabulary-in-context, author's purpose, text structure — each assessed separately.
• Comprehension question generation: Type-specific question generation per comprehension skill.

CEFR Proficiency Mapping#

CEFR (Common European Framework of Reference) is the international standard for language proficiency levels, ranging from A1 (beginner) through C2 (mastery). Mnemosyne implements the full framework plus two additional standards used in professional and academic contexts:

• Full CEFR level descriptors (A1–C2) per skill (reading, writing, speaking, listening) — what the learner can do at each level.
• ILR scale support (0–5+) with descriptors — the US government scale used for military and intelligence language assessment.
• ACTFL proficiency guidelines: The American Council on the Teaching of Foreign Languages scale (Novice–Intermediate–Advanced–Superior–Distinguished).
• Proficiency assessment per skill domain: Computes estimated CEFR level from performance evidence.

Phonetics and Pronunciation (@mnemosyne/phonetics)#

Phonetics provides the scientific foundation for understanding and practicing the sound systems of any language. The IPA database covers all major phoneme features; the higher-level modules apply those features to specific learning tasks.

• IPA (International Phonetic Alphabet) database: Phoneme definitions with place of articulation, manner of articulation, voicing, and audio examples.
• Phoneme inventory specification per language: Which sounds each language uses, and which contrasts are phonemically distinctive (i.e., change meaning).
• Minimal pairs generation: Word pairs differing by exactly one phoneme (e.g., "ship"/"sheep", "pan"/"ban") for phonemic contrast practice.
• Tonal language support: Tone system definitions for Mandarin (4 tones + neutral), Cantonese (6 tones), Vietnamese (6 tones), Thai (5 tones), and other tonal languages — including tone sandhi rules.
• Prosody modeling: Stress, rhythm, and intonation pattern definitions for connected speech.
• Extended phonetics: Allophonic variation (how phonemes vary in context without changing meaning), phonological rules (assimilation, deletion, insertion), and coarticulation effects.

Pronunciation Training (@mnemosyne/pronunciation)#

Where @mnemosyne/phonetics provides the sound-system reference data, @mnemosyne/pronunciation provides learner-facing training tools:

• Pronunciation exercise generation: Targeted drills for specific phonemic contrasts identified as problematic for a learner's L1 background.
• Phoneme-to-grapheme and grapheme-to-phoneme mappings: Bidirectional encoding-decoding rules per language, handling exceptions systematically.
• Pronunciation comparison and error detection: Acoustic comparison of learner production against target models.
• Dialect and accent variation modeling: Distinguishes legitimate accent variation from pronunciation errors — for example, not penalizing a learner for producing a British rather than American vowel.

Writing Systems and Scripts (@mnemosyne/writing)#

Writing systems vary enormously across languages — from alphabets learned in minutes to logographic systems requiring thousands of characters. This package provides script-specific pedagogy rather than a one-size-fits-all approach:

• Writing system pedagogy: Stroke order sequences for logographic scripts (Chinese, Japanese); letter formation sequences for alphabets and abjads; syllabary introduction sequences for syllabic scripts.
• Script-specific practice exercise generation: Handwriting practice, typing practice, and recognition exercises appropriate for each script type.
• Handwriting recognition integration points: Interface for connecting OCR/handwriting recognition models for real-time feedback on handwritten practice.

Classical Language Tools (@mnemosyne/classical-tools)#

Classical language learning has a distinct challenge: texts are fixed (the corpus of Latin literature will not grow), yet learners must build enormous vocabulary and grammatical knowledge before reading fluently. This package replicates and extends the Alpheios and DCC (Dickinson College Commentaries) environments for seven classical languages.

Alpheios-Style Reading Environment#

Alpheios is the gold standard for assisted reading of classical languages — clicking a word produces immediate grammatical analysis. Mnemosyne replicates and extends this model:

• Morphological pop-up analysis on click: Part of speech, case, number, gender, tense, mood, voice, stem, and lemma for any clicked word. Supports Latin, Ancient Greek, Sanskrit, Classical Arabic, Biblical Hebrew, Old Church Slavonic, and Classical Syriac.
• CEFR equivalent proficiency levels for classical languages: Maps classical language proficiency to modern equivalent levels for curriculum planning.
• Reader mode: Inline glosses and parsing assistance configurable from none (challenge mode) to full parsing (beginner support).

DCC-Style Annotated Text Platform#

The DCC model presents annotated classical texts with vocabulary and grammar helps integrated directly into the reading experience. A learner reading Caesar or Homer can hover over any word for frequency data and grammatical context:

• Annotated reader text presentation: Core vocabulary highlighted using frequency data.
• Vocabulary frequency annotation: The most frequent 1,000 words highlighted — enabling learners to prioritize high-yield vocabulary.
• Grammar help pop-ups: Contextual grammar explanations appearing at points of syntactic difficulty.
• Progress tracking through canonical texts: Marking progress through a canonical reading curriculum (Caesar's Gallic War, Vergil's Aeneid, Homer's Iliad/Odyssey, Plato's Apology, etc.).

Linguistics Analysis (@mnemosyne/linguistics)#

Linguistics provides the theoretical framework for understanding why languages work the way they do. This is useful both for advanced language learners and for instructors designing contrastive analyses:

• Phonological analysis: Phoneme inventory comparison, phonological rule application, allophonic variation mapping.
• Morphological analysis: Morpheme segmentation, paradigm identification, morphological typology classification (isolating, agglutinating, fusional, polysynthetic).
• Syntactic analysis: Constituent structure, dependency parsing, argument structure identification.
• Language typology reference: Morphological type, word order typology (SOV/SVO/VSO, etc.), case inventory, and tone system.
• Contrastive linguistics: L1–L2 structural contrast analysis generating predicted error patterns.

Classical Education: The Trivium (@mnemosyne/rhetoric)#

The trivium — grammar, logic, and rhetoric — is the foundation of classical liberal arts education. It organizes language arts from basic correctness (grammar) through argumentation (logic) to persuasion (rhetoric). This package implements all three, plus the related arts of disputation and public speaking.

Grammar (Language Arts)#

Grammar here means the formal study of English grammar in the classical tradition, not language-learning grammar:

• Parts of speech taxonomy: Noun, verb, adjective, adverb, pronoun, preposition, conjunction, interjection — with definitions and exercises.
• Sentence parsing and diagramming: Reed-Kellogg sentence diagrams, constituent analysis, and dependency representation.
• Classical trivium pedagogy: English grammar instruction aligned to Dorothy Sayers's "Lost Tools of Learning" and Charlotte Mason approaches.
• Composition instruction: Paragraph structure (topic sentence, supporting details, concluding sentence), essay structure (classical five-paragraph to complex argument), and Aristotelian argument structure.

Logic and Critical Thinking#

• Classical formal logic: Aristotelian syllogistics — categorical propositions (A/E/I/O), syllogistic figures and moods, validity testing by Venn diagram and counterexample.
• Modern formal logic: Propositional and predicate logic, truth tables, natural deduction.
• Fallacy taxonomy: The full taxonomy of informal fallacies (ad hominem, straw man, appeal to authority, false dichotomy, slippery slope, etc.) with recognition exercises and real-world examples.
• Argument mapping: Visual representation of argument structure, premise-conclusion relationships, and counter-argument positions.

Rhetoric and Persuasion#

• Aristotle's rhetorical appeals: Ethos (character and credibility), pathos (emotional appeal), logos (logical argument) — with exercises analyzing how each is deployed in model speeches and texts.
• Classical rhetorical canons: Inventio (finding arguments), dispositio (arrangement), elocutio (style), memoria (memory), actio (delivery) — the five traditional divisions of the art of rhetoric.
• Persuasive writing instruction: From claim identification through evidence selection, rebuttal, and conclusion.

Dialectic and Disputation#

• Socratic method dialogue exercises: Structured Socratic questioning sequences for any topic.
• Medieval disputatio format simulation: The formal academic disputation structure (objection, sed contra, respondeo, replies) from Scholastic philosophy.
• Thesis defense and counter-argument generation: Preparing arguments and anticipating the strongest objections.

Figures of Speech and Tropes#

The classical tradition identified and named hundreds of rhetorical figures. This package provides both a reference catalog and practice exercises:

• Comprehensive taxonomy: All major rhetorical figures — metaphor, simile, anaphora (repetition at the beginning of clauses), epistrophe (repetition at the end), chiasmus, zeugma, litotes, hyperbole, irony, metonymy, synecdoche, personification, and dozens more.
• Identification and analysis exercises: Finding figures in passages from literature, political speeches, and advertising.
• Generation of examples per figure: Generating original examples of any rhetorical figure for creative writing instruction.

Oratory and Public Speaking#

• Speech structure analysis: Identifying exordium, narratio, confirmatio, refutatio, and peroratio in model speeches.
• Delivery coaching: Pacing, emphasis, clarity, and vocal variety instruction.
• Memorization techniques for oral performance: Method of loci (the "memory palace"), the peg system, and chunking for memorizing speeches.

Art History and Visual Analysis (@mnemosyne/aesthetics)#

Art history is both a factual discipline (this painting was made in Florence in 1482) and an interpretive one (this painting means X within its iconographic tradition). The aesthetics package supports both:

• Art historical database: Artwork records with medium, period, movement, creator, location, dimensions, and iconographic content.
• Visual analysis AI: Iconographic analysis (identifying symbols, figures, and narrative content), compositional analysis (balance, rhythm, focal point), and stylometric comparison (attribution analysis).
• Iconographic analysis: Symbol identification; attribute reading (identifying saints by their symbols — St. Peter's keys, St. Catherine's wheel, St. Jerome's lion); typology classification (identifying Old Testament scenes as types prefiguring New Testament antitypes).
• Period and movement studies: Comprehensive art historical timeline from Prehistoric cave painting through Postmodern and Digital art, with style characteristics, key figures, and representative works for each period.
• Global art traditions: African, Asian, Pre-Columbian, Islamic, Oceanic, and Indigenous art studied alongside the Western canon — as independent traditions of equal depth, not peripheral exceptions.
• Architecture history: Architectural orders (Doric, Ionic, Corinthian, Tuscan, Composite); building typologies (basilica, cathedral, mosque, temple, palace); structural systems (post-and-lintel, arch, vault, dome, steel frame).
• Technical art history: Materials analysis (identifying pigments, supports, binding media); conservation science; provenance research methodology.
• Art theory and criticism: Formalist, contextual, feminist, postcolonial, Marxist, psychoanalytic, and semiotic critical frameworks — each with representative texts and application exercises.

Art media covered include oil on canvas/panel, tempera, fresco, watercolor, gouache, pastel, drawing, engraving, etching, lithography, photography, bronze and stone sculpture, wood carving, ceramic, textile, installation, video, digital, and performance art.

Mythology and Comparative Religion (@mnemosyne/mythology)#

Comparative mythology provides tools for studying religious narratives across cultures, identifying recurring motifs, and understanding myths within their religious and sociological contexts:

• Mythological database: Deity, hero, creature, and narrative records across all major world pantheons, with genealogies, epithets, domains, iconographic attributes, and major narrative roles.
• Pantheons covered: Greek, Roman, Norse, Celtic, Egyptian, Mesopotamian, Hindu, Buddhist, Chinese, Japanese, Mesoamerican (Aztec, Maya), Andean, Native American (multiple traditions), Yoruba, Yoruba diaspora (Candomblé, Santería), Polynesian, Slavic, Baltic, Finno-Ugric, and others.
• Comparative mythology analysis: Cross-cultural motif identification using the Aarne-Thompson-Uther (ATU) folktale type index; structural comparison using Vladimir Propp's narrative morphology and Claude Lévi-Strauss's structural analysis.
• Folklore and tale types: ATU index integration; international folktale classification; legend, myth, and folktale genre distinction.
• Sacred text studies: Mythological exegesis and narrative theology — interpreting myths within their religious and cultural contexts.
• Religious studies framework: Phenomenology of religion (Rudolf Otto, Mircea Eliade); comparative religious thought; ritual theory (Victor Turner, Catherine Bell); sociological approaches (Émile Durkheim, Max Weber).

Cultural Heritage Preservation (@mnemosyne/heritage)#

Digital cultural heritage preservation is both a technical problem (how do you capture and archive a three-dimensional artifact?) and a legal and ethical one (who owns the digital record of a looted object?). This package addresses all dimensions:

• 3D digitization pipeline: Photogrammetry (SfM/MVS), structured light scanning, LiDAR, CT scanning, RTI (Reflectance Transformation Imaging), multispectral imaging, and ToF (time-of-flight) methods — each appropriate for different artifact types.
• Virtual reconstruction: Evidence-based reconstruction of damaged or destroyed artifacts and sites, with explicit documentation of the evidence base and confidence levels for each reconstructed element.
• Digital archive infrastructure: Long-term preservation standards following the OAIS (Open Archival Information System) reference model; metadata schemas using Dublin Core and CIDOC CRM — the international standard for cultural heritage information modeling.
• Virtual museum builder: Interactive digital exhibitions with spatial navigation, annotation layers, multimedia content, and accessibility features.
• Conservation documentation: Condition reports (describing current artifact state), treatment records (documenting conservation interventions), and environmental monitoring (tracking temperature, humidity, light exposure).
• Provenance and repatriation tracking: Ownership chain documentation from creation through all transfers of custody; legal status tracking; repatriation claim management in compliance with UNESCO conventions.
• Intangible heritage documentation: Oral traditions, performing arts, craftsmanship, ritual practices, and festive events, following the UNESCO 2003 Convention on Intangible Cultural Heritage framework.

History, Archaeology, and Anthropology (@mnemosyne/temporal)#

The temporal package treats historical knowledge as a structured dataset: events have causes and consequences, people exist in social networks, sites have stratigraphic layers, and dates need calendar-system context.

• Historical database infrastructure: Event, person, place, and period records with multi-calendar support — Gregorian, Julian, Coptic, Islamic (Hijri), Hebrew, Chinese, Mayan (Long Count), Roman (AUC), and Egyptian calendars, all with bidirectional conversion.
• Prosopographical tools: Person records, biographical data, and relationship networks across historical populations. Prosopography is the discipline of studying populations through systematic compilation of individual records, essential for understanding social networks in ancient societies.
• Geographic historical analysis: Territorial mapping over time (the same city under multiple political entities); migration route analysis (Silk Road, Bantu migrations, Indo-European dispersal); trade network reconstruction.
• Archaeological intelligence: Excavation site records with stratigraphic context (the sequential layers of occupation that provide relative dating); artifact classification following standard typological schemas; site GIS integration.
• AI archaeological discovery: Pattern recognition in aerial survey data and LiDAR scans; remote sensing analysis for identifying subsurface features.
• Anthropological frameworks: Cultural evolution models; kinship system analysis (unilineal, bilateral, cognatic descent); social organization analysis (band, tribe, chiefdom, state progression); ethnographic fieldwork methodology.
• Bioarchaeology and human origins: Skeletal analysis (age-at-death, sex, pathology estimation); paleopathology (disease in archaeological populations); ancient DNA interpretation; stable isotope analysis (revealing diet and migration patterns from bone chemistry).
• Economic history analysis: Price history databases; commodity network reconstruction; monetary system evolution; long-term economic trend analysis.
• Military and political history: Battle analysis (terrain, tactics, casualties, outcome factors); state formation models; political succession tracking.

Immersive Language Learning (@mnemosyne/immersion)#

Immersion methodology is based on Stephen Krashen's Input Hypothesis — that language is acquired (not explicitly learned) through comprehensible input at the i+1 level (slightly beyond current ability) — and its modern extensions through the Refold and AJATT/MIA communities.

Comprehensible Input Engine#

The core challenge of immersion is finding material at exactly the right difficulty level:

• i+1 difficulty scoring: Automatic scoring of any text or video segment for comprehensibility at the learner's current level. Material fully understood (i+0) is not challenging; material with too many unknowns (i+2+) produces anxiety rather than acquisition; i+1 is the optimal acquisition zone.
• MorphMan-style morpheme frequency analysis: Ranks text items by the morphemes they contain, so the most frequent and therefore most acquisition-valuable items are presented first.
• Known word tracking: Persistent tracking of all words encountered and their status across every content source.
• Refold-style stage progression system: Structured progression through stages from total beginner (stage 1, structured study) through upper beginner immersion (stage 2) to intermediate and advanced immersion stages — each with appropriate content recommendations and study activities.
• Acquisition vs. learning mode distinction: analyseAcquisitionBalance distinguishes formal study (learning) from naturalistic exposure (acquisition) and tracks both separately.

Sentence Mining System#

Sentence mining is the practice of extracting sentences from native content to create personalized SRS study cards — combining the benefits of immersion with the efficiency of spaced repetition:

• Subtitle parsing and sentence extraction: parseSRT parses SRT subtitle files; extractMiningSentences extracts subtitle-aligned candidate sentences as study material.
• 1T (one-target) sentence identification: filterOneTargetSentences finds sentences where exactly one word is unknown — the optimal difficulty for vocabulary acquisition from context.
• Sentence quality and deduplication: scoreSentenceQuality, computeJaccardSentenceSimilarity, and detectNearDuplicates rank mined sentences and remove near-duplicates.
• Bilingual subtitle alignment: alignBilingualSubtitles pairs L1 and L2 subtitle tracks for parallel comprehension.
• Card templates: CARD_TEMPLATES defines sentence, vocabulary, audio-only, picture-sentence, and cloze card layouts for mined material.
• Media capture (not implemented): Screenshot/audio clip capture, GIF generation, and Whisper-based transcription of audio without subtitles are not present in the current library.

Video Immersion Platform#

• Graded video library: VIDEO_DIFFICULTY_TIERS, VideoContent, CreatorProfile, and classifyVideoDifficulty organize content by difficulty with creator/accent metadata.
• Interactive subtitles: buildInteractiveSubtitle produces clickable, word-level subtitles (SubtitleWord / InteractiveSubtitle); watch history is recorded via VideoWatchHistoryRecord.
• Streaming-platform enhancement (not implemented): A Netflix/YouTube browser extension is not part of this library.

Reading Immersion Tools#

• LingQ-style word tracking: TrackedWord, createTrackedWord, and advanceWordStatus track per-word status that persists across reading sessions; ReadingSessionRecord / computeReadingStats accumulate progress.
• Word familiarity levels (1–5 scale): WordStatus (1|2|3|4|5|'known'|'ignored') and WORD_FAMILIARITY_LEVELS track a word from first encounter through incidental recognition to active production.
• Graded readers: GRADED_READER_CATALOG (GradedReader) provides a difficulty-tiered reader catalog.
• Parallel text reader: buildParallelText (ParallelTextSegment) places original and translation side by side.
• Popup dictionary: buildPopupEntry (PopupDictionaryEntry) supplies in-reader word lookups. A full epub/PDF document reader is not implemented.

Listening Immersion Tools#

• Podcast integration: PodcastFeed / PodcastEpisode, SAMPLE_COMPREHENSIBLE_PODCASTS, recommendPodcasts, and a listening journal (ListeningJournalEntry) support transcript-backed listening practice.
• Listening assessment and exercises: assessListeningLevel, generateGapFillExercise, and generateComprehensionQuiz produce graded listening activities; RadioStation / SAMPLE_RADIO_STATIONS add live audio.
• Condensed audio: AudioCondensationConfig / DEFAULT_CONDENSATION_CONFIG configure silence-removal for listening efficiency.
• Shadowing (not implemented): Automatic shadowing-exercise generation is not present in the current library.

Advanced Gamification (@mnemosyne/gamification-plus)#

This package layers a Duolingo-style competitive and social system on top of the core gamification in @mnemosyne/experience. The two packages are independent: gamification-plus adds no code dependency on experience.

• League system: Six-tier competitive leagues — Bronze, Silver, Gold, Platinum, Diamond, Obsidian — with weekly promotion (top N in league move up) and demotion (bottom N move down) zones, creating ongoing competitive motivation.
• Streak system: Daily streak tracking with streak freeze (protecting the streak if a day is missed, purchased with tokens) and streak repair (restoring a broken streak within a short window).
• Weekly goals: Configurable XP targets with safe-zone thresholds that protect against demotion, removing the anxiety of falling behind mid-week.
• Seasonal events: Time-limited events with bonus XP multipliers, exclusive cosmetic rewards, and narrative framing tied to seasons or cultural moments.
• Double-XP weekends: Scheduled bonus periods to drive re-engagement.
• Collaborative community goals: Group challenges requiring collective contribution from all members of a study group or community, building social learning bonds.
• Token economy: League reward tokens spendable on streak freezes, cosmetic customization, XP boosts, and other rewards.

Learning Experience and Adaptive Engine (@mnemosyne/experience)#

Knowledge Tracing Models#

Where @mnemosyne/core provides static retention models (Ebbinghaus, SM-2, FSRS), @mnemosyne/experience provides dynamic skill-state models that update as the learner practices:

• Bayesian Knowledge Tracing (BKT): A 4-parameter model per skill (BKTParams, DEFAULT_BKT_PARAMS), tracking the probability that a learner has mastered each skill and updating after each response via updateBKT. The four parameters are: p(initial mastery), p(learning from practice), p(slip — wrong despite mastery), and p(guess — right despite non-mastery).
• Deep Knowledge Tracing (DKT): An LSTM-style model (DKTState, updateDKTState) that decays a hidden skill-state representation across responses.
• Adaptive selection: Multi-armed bandit (ucb1SelectArm) and a reinforcement-learning policy (selectRLAction) drive content selection. classifyZPD and optimiseLearningTrajectory keep learners in the Zone of Proximal Development.

Progress and Social Features#

• Gamification system: XP and levels (awardXP, LEVEL_DEFINITIONS), achievements (ACHIEVEMENT_CATALOG, checkAchievements), daily challenges, leaderboards, a virtual-currency shop, and avatar customization.
• Social learning: Study groups (createStudyGroup, joinStudyGroup), team challenges (updateTeamChallengeProgress), tutor/tutee matching (matchTutorTutee), and shareable progress cards.
• Skill trees and quests: HUMANITIES_SKILL_TREE and SAMPLE_QUESTS structure long-form progression; issueCertificate issues completion certificates.
• Accessibility and inclusion: WCAG criteria auditing (auditWCAGCompliance, WCAG_AAA_CRITERIA), ARIA configuration, keyboard-shortcut maps, color-vision-adapted palettes (COLOR_PALETTES), typography presets (dyslexia-friendly and others), focus mode, break reminders, and contrast-ratio checking.

Platform Integration (@mnemosyne/platform)#

@mnemosyne/platform solves the interoperability problem: learners carry their study data across tools (Anki, university LMS systems, Zotero, museum APIs), and this package provides the parsing and URL-building layer that makes those transfers possible without network I/O inside the domain itself.

• Anki import/export: parseAnkiDeck / exportAnkiDeck parse and generate Anki decks — notes, note models, and subdeck hierarchy — from a simplified Anki deck JSON export (an .apkg deconstruction, not the binary archive).
• CSV vocabulary import: parseCSVVocabulary ingests vocabulary lists.
• LMS / e-learning standards: SCORM manifest parsing (parseSCORMManifest), xAPI statement creation (createXAPIStatement, XAPI_VERBS), and LTI launch validation (validateLTILaunch).
• Reference and portability: Zotero export parsing (parseZoteroExport), portable progress data (buildPortableProgressData), and GDPR data-package assembly (buildGDPRPackage).
• External content connectors: URL builders and payload parsers for Wikipedia/Wikidata, Europeana, Internet Archive, dictionary providers, translation providers, museum APIs, and SRU library catalogs.
• API descriptors: APIEndpointDefinition and PUBLIC_API_ENDPOINTS describe a REST surface a hosting application could expose — declarative metadata, not a running server.

The package does not implement binary .apkg parsing, EPUB/PDF ingestion, TTS/STT, multi-platform runtimes, or a versioned content-management workflow.

Community Features (@mnemosyne/community)#

Language Exchange#

Language exchange connects learners who each speak the other's target language, enabling bilateral practice that neither can get from solo study:

• Partner matching algorithm: computeMatchScore and findLanguagePartners match learners factoring in interests, timezone, proficiency level, and goals (ExchangeProfile, LanguageProfile, MatchScore).
• Exchange sessions: scheduleExchangeSession and checkTimeSplitBalance schedule and balance bilateral practice sessions; PartnerRelationship tracks ongoing partnerships.
• Text chat with inline correction tools: applyInlineCorrection and renderCorrectionMarkup render corrections as tracked changes; saveVocabularyFromCorrection turns corrections into vocabulary items; suggestConversationTopics (TOPIC_SUGGESTIONS) seeds conversations.
• Voice rooms (drop-in conversation practice): createVoiceRoom, joinVoiceRoom, and toggleHandRaise run topic-based audio rooms with live transcript segments and participation stats.
• Social moments feed: createMoment, addMomentCorrection, and filterMomentsForLearner provide short target-language posts that others can correct inline; computeMomentAnalytics and detectSpam support moderation.
• Shared whiteboard: createWhiteboard / addWhiteboardElement provide a collaborative whiteboard scoped to an exchange session.
• Transliteration: A transliterate dispatcher with per-script implementations (kana, kanji, Hangul, Cyrillic, Arabic, Devanagari, Thai, Georgian) helps partners read each other's scripts.
• Rating and reporting: createPartnerRating / computeAverageRating and createUserReport support partner feedback and safety.

A dedicated native-speaker feedback marketplace, community deck sharing, discussion forums, and expert-contributor recognition are not implemented in the @mnemosyne/community package. Collaborative text annotation with upvoting and editing is implemented, but in @mnemosyne/classical-tools.

Philological Tools (@mnemosyne/philology)#

Philology is the study of language in written historical sources. Where @mnemosyne/linguistics covers synchronic (present-state) analysis, @mnemosyne/philology covers diachronic (historical change) analysis and textual criticism:

• Etymology tracing: Word origin and cognate network visualization — tracing the history of a word from its reconstructed proto-language root through all its descendant forms.
• Semantic change analysis: Diachronic documentation of semantic shifts — how word meanings have widened, narrowed, ameliorated, or pejorized across time.
• Stylometric analysis: Authorship attribution tools using function word frequencies, sentence complexity metrics, and vocabulary profile statistics.
• Corpus frequency analysis: Word frequency distributions across texts; comparative frequency between corpora (e.g., a word common in Classical Latin but rare in Medieval Latin).
• Collocational analysis: Documents typical word combinations and usage patterns — essential for natural-sounding production.

Scope and Domain Boundary#

This feature document is scoped to libs/mnemosyne/* (Phase 39). It covers all 19 packages: core, polyglot, phonetics, linguistics, classical-tools, philology, mythology, aesthetics, rhetoric, heritage, temporal, knowledge-graph, immersion, experience, gamification-plus, community, pronunciation, writing, and platform. Items marked "(not implemented)" are described because they appear in the Phase 39 vision but have no corresponding code in the current library.

Nisaba owns ancient-text scholarly analysis and manuscript/corpus research. Mnemosyne owns cultural learning, language acquisition, pronunciation, humanities education, and cultural heritage experiences. The distinction is that Mnemosyne is pedagogical — its goal is a learner acquiring knowledge — while Nisaba is research-oriented, concerned with the scholarly study of texts as primary sources.

Autonomous Research Contribution (Phase 178)#

Mnemosyne is a co-owner of the autonomous research / agentic-scientist substrate (Phase 178, centered in Nous). Mnemosyne's side is humanistic and cultural-knowledge grounding plus long-horizon memory: the agentic scientist draws on Mnemosyne's knowledge graph and spaced-memory surfaces for research that spans languages and traditions, and contributes back to the shared experiment/knowledge ledger. Nous owns the agent loop; Mnemosyne owns the cultural-knowledge model.