OMS for Education and Research: Knowledge Graphs, Citations, and Reproducibility

Educational AI agents face a challenge that sets them apart from most other agent types: they need to build structured knowledge incrementally, track a learner's progress over time, maintain rigorous citation trails back to source materials, and — in research settings — ensure that every conclusion is reproducible from the original data. These are not nice-to-have features. They are fundamental requirements for any system trusted with teaching and scholarship.

The Open Memory Specification addresses each of these requirements through its core grain types and cross-linking mechanisms. This post maps OMS to education and research, showing how Belief grains form knowledge graph triples, provenance chains serve as citation trails, Events capture student interactions, Goals track learning objectives, and embedding references enable semantic search across accumulated knowledge.

Beliefs as knowledge graph triples

The Belief grain type (Section 8.1) is a structured knowledge claim modeled as a semantic triple: subject-relation-object. This maps directly to the nodes and edges of a knowledge graph.

Consider the complete example grain from Appendix F of the spec:

{
  "type": "belief",
  "subject": "machine-learning",
  "relation": "is_subset_of",
  "object": "artificial-intelligence",
  "confidence": 0.99,
  "source_type": "user_explicit",
  "created_at": 1737000000000,
  "namespace": "knowledge-base",
  "author_did": "did:key:z6MkhaXgBZDvotDkL5257faiztiGiC2QtKLGpbnnEGta2doK",
  "user_id": "researcher-alice",
  "importance": 0.95,
  "structural_tags": ["ai", "ml", "education"],
  "context": {"source": "textbook", "chapter": "1.2"},
  "provenance_chain": [
    {"source_hash": "abc123...", "method": "direct_input", "weight": 1.0}
  ],
  "related_to": [
    {
      "hash": "def456...",
      "relation_type": "elaborates",
      "weight": 0.8
    }
  ]
}

This grain establishes that "machine-learning is_subset_of artificial-intelligence" with 0.99 confidence, sourced explicitly from a user (source_type: "user_explicit"). The context map carries source attribution: {"source": "textbook", "chapter": "1.2"}. The structural_tags classify it under ["ai", "ml", "education"]. The related_to cross-link points to another grain that elaborates on this relationship.

Building a course knowledge graph

A course knowledge graph is built by accumulating Belief grains where concepts serve as subjects and objects, and relationships serve as relations. The subject-relation-object model from Section 8.1 maps naturally to knowledge graph nodes and edges:

{
  "type": "belief",
  "subject": "linear-algebra",
  "relation": "prerequisite_for",
  "object": "machine-learning",
  "confidence": 0.95,
  "source_type": "user_explicit",
  "created_at": 1768471200000,
  "namespace": "course:cs101:fall2026",
  "structural_tags": ["math", "prerequisites", "curriculum"],
  "context": {"source": "syllabus", "section": "prerequisites"}
}

{
  "type": "belief",
  "subject": "gradient-descent",
  "relation": "is_subset_of",
  "object": "optimization-algorithms",
  "confidence": 0.98,
  "source_type": "user_explicit",
  "created_at": 1768471200100,
  "namespace": "course:cs101:fall2026",
  "structural_tags": ["ml", "optimization", "algorithms"]
}

{
  "type": "belief",
  "subject": "neural-networks",
  "relation": "related_to",
  "object": "gradient-descent",
  "confidence": 0.92,
  "source_type": "consolidated",
  "created_at": 1768471200200,
  "namespace": "course:cs101:fall2026",
  "structural_tags": ["ml", "neural-networks", "training"]
}

Each Belief grain is a knowledge graph edge. The subject and object are nodes. The relation is the edge label. The confidence field (float64, range [0.0, 1.0] per Section 8.1) indicates how certain we are about the relationship. The structural_tags field (Section 6.1) provides topic classification for filtering and browsing.

The OMS spec notes in Section 8.1 that the Fact type follows RDF mapping: <grain:subject> <grain:relation> "grain:object" . This means any OMS knowledge graph can be exported to standard RDF/SPARQL systems for interoperability with existing educational knowledge bases.

Provenance chains as citation trails

Research integrity depends on citation — every claim must trace back to its sources. OMS builds this into the grain format through two complementary mechanisms: the provenance_chain field (Section 14.1) and the related_to cross-links (Section 14.2).

From source to conclusion

Every derived Fact carries a provenance_chain — an array of entries, each with a source_hash (content address of the source grain), method (how it was derived), and weight (how much this source contributed, range 0.0 to 1.0):

{
  "type": "belief",
  "subject": "spaced-repetition",
  "relation": "improves",
  "object": "long-term-retention",
  "confidence": 0.88,
  "source_type": "consolidated",
  "created_at": 1768471200000,
  "namespace": "research:lab:nlp",
  "provenance_chain": [
    {
      "source_hash": "<hash-of-study-1-episode>",
      "method": "direct_input",
      "weight": 0.6
    },
    {
      "source_hash": "<hash-of-study-2-episode>",
      "method": "frequency_consolidation",
      "weight": 0.4
    }
  ],
  "derived_from": [
    "<hash-of-study-1-episode>",
    "<hash-of-study-2-episode>"
  ]
}

The provenance_chain records the full derivation trail. The derived_from field (Section 6.1) carries the parent content addresses. Together, they create a citation graph that traces every conclusion back to its source data.

Cross-links for scholarly relationships

The related_to field (Section 14.2) enables semantic links between grains using a closed vocabulary of relation types (Section 14.3). Several of these map directly to scholarly relationships:

A research finding that is supported by multiple independent studies can express this through cross-links:

{
  "type": "belief",
  "subject": "transformer-attention",
  "relation": "outperforms",
  "object": "recurrent-models-on-long-sequences",
  "confidence": 0.93,
  "source_type": "consolidated",
  "created_at": 1768471200000,
  "namespace": "research:lab:nlp",
  "related_to": [
    {
      "hash": "<hash-of-empirical-study-1>",
      "relation_type": "supports",
      "weight": 0.9
    },
    {
      "hash": "<hash-of-empirical-study-2>",
      "relation_type": "supports",
      "weight": 0.85
    },
    {
      "hash": "<hash-of-theoretical-analysis>",
      "relation_type": "elaborates",
      "weight": 0.7
    },
    {
      "hash": "<hash-of-general-architecture-claim>",
      "relation_type": "generalizes",
      "weight": 0.6
    }
  ]
}

Each cross-link entry uses field compaction (Section 14.2): hash becomes h, relation_type becomes rl, and weight becomes w. The weight field is a float64 indicating how strongly this link applies.

Episodes for student interactions

Every tutoring conversation, question-answer session, and lab notebook entry is raw interaction data — the kind of unstructured record that the Episode memory type (Section 8.2) is designed to capture.

Capturing interactions

{
  "type": "event",
  "content": "Student asked: 'Why does gradient descent sometimes get stuck in local minima?' Tutor explained saddle points, learning rate schedules, and the difference between convex and non-convex optimization. Student followed up with a question about Adam optimizer.",
  "created_at": 1768471200000,
  "user_id": "student-bob-042",
  "namespace": "course:cs101:fall2026",
  "importance": 0.6,
  "consolidated": false,
  "structural_tags": ["tutoring", "optimization", "gradient-descent"]
}

The consolidated field (Section 8.2) starts as false — this Episode has not yet been processed. Once a consolidation process extracts structured Facts from this interaction, the Episode is marked as consolidated. Since grains are immutable, "marking" means creating a new Episode grain with consolidated: true that supersedes the original through the supersession chain.

The consolidation pipeline (described in the spec's Episode lifecycle) works like this:

1. Capture: Raw interaction recorded as Episode with consolidated absent (defaults to false)
2. Consolidation: An LLM or pattern-matching engine extracts structured Facts
3. Tracking: Extracted Facts carry consolidation_level (0=raw, 1=frequency, 2=pattern, 3=sequence)
4. Provenance: Facts link back to source Episodes via derived_from and provenance_chain
5. Marking: Original Episode superseded by a new grain with consolidated: true

From the tutoring session above, consolidation might extract:

{
  "type": "belief",
  "subject": "student-bob-042",
  "relation": "has_difficulty_with",
  "object": "local-minima-in-gradient-descent",
  "confidence": 0.8,
  "source_type": "consolidated",
  "created_at": 1768471200500,
  "consolidation_level": 0,
  "derived_from": ["<hash-of-tutoring-episode>"],
  "namespace": "student:bob:portfolio",
  "provenance_chain": [
    {
      "source_hash": "<hash-of-tutoring-episode>",
      "method": "direct_input",
      "weight": 1.0
    }
  ]
}

Multimedia episodes with content references

Lab sessions and lecture recordings are not just text. The content_refs field (Section 7.1) allows Episodes to reference multimedia content:

{
  "type": "event",
  "content": "Week 5 lecture on convolutional neural networks. Covered kernel operations, pooling layers, and feature hierarchies. Live coding demo of LeNet-5 implementation.",
  "created_at": 1768471200000,
  "namespace": "course:cs101:fall2026",
  "structural_tags": ["lecture", "cnn", "week-5"],
  "content_refs": [
    {
      "uri": "cas://sha256:a1b2c3d4...",
      "modality": "video",
      "mime_type": "video/mp4",
      "size_bytes": 524288000,
      "checksum": "sha256:a1b2c3d4...",
      "metadata": {"width": 1920, "height": 1080, "fps": 30, "duration_ms": 3600000, "codec": "h264"}
    },
    {
      "uri": "cas://sha256:b2c3d4e5...",
      "modality": "audio",
      "mime_type": "audio/aac",
      "size_bytes": 28800000,
      "checksum": "sha256:b2c3d4e5...",
      "metadata": {"sample_rate_hz": 48000, "channels": 2, "duration_ms": 3600000}
    },
    {
      "uri": "cas://sha256:c3d4e5f6...",
      "modality": "image",
      "mime_type": "image/png",
      "size_bytes": 2048576,
      "checksum": "sha256:c3d4e5f6...",
      "metadata": {"width": 1920, "height": 1080, "color_space": "sRGB"}
    }
  ]
}

The video metadata schema (Section 7.3) includes width, height, fps, duration_ms, and codec. The audio schema includes sample_rate_hz, channels, and duration_ms. The image schema includes width, height, and color_space. Each content reference carries a checksum for integrity verification — Section 20.5 requires that implementations verify checksums after fetching.

Goals for learning objectives

The Goal memory type (Section 8.7) maps directly to learning objectives with lifecycle semantics that track a student's progress from enrollment through mastery.

Curriculum hierarchy

Goals support DAG-structured hierarchies through the parent_goals field — an array of content addresses, not a single parent pointer. This enables curriculum modeling:

Course-level goal:

{
  "type": "goal",
  "subject": "student-alice-017",
  "description": "Master linear algebra fundamentals",
  "goal_state": "active",
  "source_type": "user_explicit",
  "created_at": 1768471200000,
  "namespace": "course:cs101:fall2026",
  "criteria": [
    "solve systems of equations",
    "compute eigenvalues",
    "understand vector spaces"
  ],
  "criteria_structured": [
    {
      "metric": "assessment_score_linear_systems",
      "operator": "gte",
      "threshold": 0.8,
      "measurement_ns": "course:cs101:assessments"
    },
    {
      "metric": "assessment_score_eigenvalues",
      "operator": "gte",
      "threshold": 0.75,
      "measurement_ns": "course:cs101:assessments"
    }
  ],
  "priority": 2,
  "progress": 0.0
}

Module-level goal:

{
  "type": "goal",
  "subject": "student-alice-017",
  "description": "Complete systems of equations module",
  "goal_state": "active",
  "source_type": "system",
  "created_at": 1768471200100,
  "namespace": "course:cs101:fall2026",
  "parent_goals": ["<hash-of-master-linear-algebra-goal>"],
  "criteria": ["solve 2x2 systems", "solve 3x3 systems", "apply Gaussian elimination"],
  "priority": 2,
  "progress": 0.0,
  "provenance_chain": [
    {
      "source_hash": "<hash-of-master-linear-algebra-goal>",
      "method": "goal_decomposition",
      "weight": 1.0
    }
  ]
}

Lesson-level goal:

{
  "type": "goal",
  "subject": "student-alice-017",
  "description": "Solve 2x2 linear systems using substitution and elimination",
  "goal_state": "active",
  "source_type": "system",
  "created_at": 1768471200200,
  "namespace": "course:cs101:fall2026",
  "parent_goals": ["<hash-of-systems-of-equations-goal>"],
  "criteria_structured": [
    {
      "metric": "practice_problems_correct_ratio",
      "operator": "gte",
      "threshold": 0.85,
      "window_ms": 604800000,
      "measurement_ns": "course:cs101:practice"
    }
  ],
  "priority": 3,
  "progress": 0.0
}

The hierarchy flows: course goal decomposes to module goals, module goals decompose to lesson goals. Each uses method: "goal_decomposition" in the provenance chain (Section 8.7, provenance chain methods table).

Progress tracking and state lifecycle

The progress field (Section 8.7) is a float64 in range [0.0, 1.0] — an agent-assessed progress estimate. As a student completes practice problems and assessments, the agent updates progress by creating new Goal grains through the supersession chain:

G1: goal_state="active", progress=0.0
G2: goal_state="active", progress=0.45, derived_from=[<hash-G1>]
G3: goal_state="active", progress=0.82, derived_from=[<hash-G2>]
G4: goal_state="satisfied", progress=1.0, derived_from=[<hash-G3>],
    satisfaction_evidence=[<assessment-toolcall-hash>, <practice-observation-hash>],
    state_reason="All criteria verified — assessment score 0.88 exceeds threshold 0.8"

The goal_state lifecycle has four states (Section 8.7): "active", "satisfied", "failed", and "suspended". A course break maps to "suspended". A completed module maps to "satisfied" with satisfaction_evidence referencing the assessment ToolCall or Observation grains that substantiate the transition. Each state transition creates a new immutable grain in the supersession chain.

Embedding references for semantic search

As a knowledge graph grows — hundreds of concepts, thousands of relationships — students and researchers need to find relevant information by meaning, not just by keyword. OMS supports this through embedding references (Section 7.2).

Each grain can carry an embedding_refs array:

{
  "type": "belief",
  "subject": "backpropagation",
  "relation": "is_algorithm_for",
  "object": "training-neural-networks",
  "confidence": 0.99,
  "source_type": "user_explicit",
  "created_at": 1768471200000,
  "namespace": "course:cs101:fall2026",
  "embedding_refs": [
    {
      "vector_id": "vec-bp-001",
      "model": "text-embedding-3-large",
      "dimensions": 3072,
      "modality_source": "text",
      "distance_metric": "cosine"
    }
  ]
}

The embedding reference schema (Section 7.2) has three required fields: vector_id (ID in the vector store), model (embedding model name), and dimensions (vector dimensionality, e.g., 3072 for text-embedding-3-large). The optional modality_source field indicates what was embedded ("text", "image", "audio", etc.), and distance_metric specifies the comparison metric ("cosine", "l2", "dot").

With embedding refs across the knowledge graph, a student can query "find concepts related to gradient computation" and get semantically similar grains — backpropagation, chain rule, automatic differentiation — ranked by cosine similarity. The vector search happens in the external vector store; the OMS grain carries the reference that links the knowledge graph node to its vector representation.

The has_embedding_refs flag (bit 4 in the header flags byte, Section 3.1.1) enables O(1) filtering: scan grain headers to find all grains with associated embeddings without deserializing any payloads.

Research reproducibility

Research reproducibility is perhaps the most compelling application of OMS in academic settings. The core requirement is that every conclusion traces back through a verifiable chain to the original data and methods.

The reproducibility chain

OMS achieves this through the convergence of three features:

Content addressing (Section 5) ensures that every grain — every data point, every method call, every intermediate result, every conclusion — is identified by the SHA-256 hash of its exact binary representation. If a byte changes, the hash changes. This means "grain X" refers to exactly one immutable set of bytes, forever.

Provenance chains (Section 14.1) trace every derived grain back to its sources:

Conclusion (Fact):
  "spaced-repetition improves long-term-retention"
  confidence: 0.88
  provenance_chain:
    → source_hash: <analysis-toolcall>, method: "frequency_consolidation"
    → source_hash: <study-1-episode>, method: "direct_input"

Analysis (ToolCall):
  tool_name: "statistical_analysis"
  arguments: {"method": "paired_t_test", "alpha": 0.05}
  result: {"t_statistic": 4.23, "p_value": 0.0002}
  success: true
  derived_from: [<study-1-episode>, <study-2-episode>]

Study 1 (Episode):
  content: "Experiment session 1: 40 participants..."
  content_refs: [{modality: "document", uri: "cas://sha256:..."}]

Study 2 (Episode):
  content: "Experiment session 2: 38 participants..."
  content_refs: [{modality: "document", uri: "cas://sha256:..."}]

Every conclusion (Belief) traces through the provenance chain to the analysis methods (Actions) and source data (Events, Observations). The Action grain records exactly what analysis was run (tool_name, input), whether it errored (is_error), and what the result was (content). Content references on the Event grains point to the raw experimental data with SHA-256 checksums for integrity.

Immutability ensures that none of these grains can be modified after creation. The original data, the analysis method, and the conclusion are all permanent records. If the analysis needs to be re-run with different parameters, a new Action grain is created — it does not replace the old one. If the conclusion changes, a new Belief grain supersedes the previous one through the supersession chain. The complete history is always available.

Actions as method records

The Action grain type (Section 8.5) is particularly valuable for research reproducibility. Every computational step — statistical tests, model training runs, data transformations — is recorded as an Action grain:

{
  "type": "action",
  "tool_name": "train_classifier",
  "input": {
    "model": "logistic_regression",
    "features": ["study_hours", "practice_score", "attendance"],
    "target": "exam_pass",
    "train_split": 0.8,
    "random_seed": 42
  },
  "content": {
    "accuracy": 0.87,
    "precision": 0.85,
    "recall": 0.89,
    "f1": 0.87
  },
  "is_error": false,
  "duration_ms": 4500,
  "created_at": 1768471200000,
  "namespace": "research:lab:nlp",
  "author_did": "did:key:z6MkResearcherAlice..."
}

The arguments map records every parameter. The result field records every output metric. The duration_ms field records execution time. Another researcher can verify the exact method used and either reproduce it or identify precisely where their methodology differs.

Namespace strategy

OMS namespaces (Section 6.1, namespace field, default "shared") partition memory into logical spaces. For education and research, a hierarchical naming convention keeps content organized:

The first two bytes of SHA-256(namespace) are stored in the fixed header (bytes 3-4) as a routing hint (Section 3.1.1), enabling efficient namespace-based filtering without deserialization. A query for "all grains in course:cs101:fall2026" can pre-filter by namespace hash before any payload parsing.

Namespace separation also has compliance implications. A student's personal data lives in student:alice:portfolio with the user_id field set for GDPR scoping. Course-level content in course:cs101:fall2026 is organizational. Research data in research:lab:nlp may carry different sensitivity classifications (Section 13.1) depending on whether it involves human subjects.

The complete picture: a semester of learning

Here is how the pieces fit together across a semester.

Week 1: The course knowledge graph is seeded with Belief grains — concept definitions, prerequisite relationships, topic classifications. Each grain carries structural_tags for topic navigation and context maps for source attribution (textbook, chapter, section).

Weeks 2-14: As students interact with the AI tutor, Episodes capture every tutoring session, question-answer exchange, and lab notebook entry. The consolidated field tracks which episodes have been processed. Consolidation extracts Facts about student understanding — what they know, what they struggle with, what misconceptions they hold.

Throughout: Goal grains track learning objectives at course, module, and lesson levels. Progress updates flow through the supersession chain. Assessment results become Action grains and Observation grains that serve as satisfaction_evidence for goal completion.

Research component: Students conducting research have their experimental data captured as Events with content references to raw data files. Analysis steps are recorded as Action grains. Conclusions become Beliefs with provenance chains tracing back through every analysis to the original data.

Semantic search: Embedding references across the knowledge graph enable "find concepts similar to X" queries, helping students discover connections they might not have found through keyword search alone.

At semester end: The complete learning history — every interaction, every assessment, every knowledge claim, every citation chain — is a collection of immutable, content-addressed grains. A student's portfolio in namespace student:alice:portfolio is portable, verifiable, and auditable. A researcher's findings in namespace research:lab:nlp are reproducible by anyone with access to the grain collection and the referenced data.

Summary

OMS maps to education and research through the natural alignment between its memory types and academic needs. Facts model knowledge graph triples with subject-relation-object semantics, confidence scores, and source attribution. Provenance chains serve as citation trails where every derived claim traces back to source materials through content-addressed links. Episodes capture the raw interactions — tutoring sessions, lab notes, lectures — that feed the consolidation pipeline. Goals track learning objectives through a lifecycle of active, satisfied, failed, and suspended states, with machine-evaluable criteria and DAG-structured curriculum hierarchies. Embedding references enable semantic search across the entire knowledge graph. And content addressing guarantees that every grain in a research chain is immutable and verifiable — the foundation of reproducible scholarship.

The result is a memory format where knowledge graphs are built from verifiable triples, every conclusion cites its sources through cryptographic links, every student interaction is preserved for longitudinal analysis, and every research finding is reproducible from first principles.