Cognitive Triad Taxonomy (Semantic / Episodic / Procedural)

ADR-001: Cognitive Triad Taxonomy (Semantic / Episodic / Procedural)

Status

Accepted

Context

Background and Problem Statement

MIF needs a memory classification system that anchors every concept in a small, stable set of base types. That taxonomy has to satisfy four constraints at once: it must be universally applicable across domains, grounded in established theory rather than ad-hoc opinion, simple enough to be used without specialist training, and expressive enough to organize the memory of AI systems that accumulate facts, events, and procedures over time.

This requirement is sharpened by MIF’s relationship to the Open Knowledge Format (OKF). OKF defines a minimal interoperability surface and deliberately refuses to define a concept-type taxonomy — it requires a type field on every concept but says nothing about what the values should be. MIF’s positioning thesis (SPECIFICATION.md, “MIF answers OKF’s open questions”) is to supply opinionated answers to exactly those open questions. The base concept-type taxonomy is the first such answer: it is MIF’s response to OKF’s deliberately-absent concept-type system.

Current Limitations

Without a principled taxonomy, a memory format drifts toward one of several failure modes:

• Fragmentation: Domain-specific category sets do not port between systems; a memory typed for one application is meaningless to another.
• No semantic hierarchy: Flat tagging records keywords but encodes no organizing structure, so consumers cannot reason about what kind of thing a memory is.
• No interoperability: A custom ontology invented per system cannot satisfy OKF’s requirement that every bundle be readable by a generic consumer.
• Unverifiable foundation: A taxonomy invented from intuition has no external authority; reviewers cannot evaluate whether the categories are complete or well-separated.

Decision Drivers

Primary Decision Drivers

1. Universality: The base types must apply to any domain — engineering, operations, personal knowledge — without per-domain forks.
2. Theoretical grounding: The categories must rest on established research, not ad-hoc invention, so the boundaries between them are defensible.
3. Answering OKF’s open concept-type question: MIF must fill the taxonomy OKF leaves blank while keeping each typed concept legible to a generic OKF reader.

Secondary Decision Drivers

1. Simplicity: A small, memorable set of types lowers the cost of adoption for both human authors and AI systems.
2. Extensibility: The base set must remain small while still allowing domain-specific specialization without breaking interoperability.

Considered Options

Option 1: Domain-specific categories

Description: Define category sets tailored to each application domain (e.g., “bug”, “feature”, “meeting” for engineering; different sets elsewhere).

Advantages:

• Maximally expressive within a single domain.
• Categories map directly onto familiar local vocabulary.

Disadvantages:

• Fragmented: no shared taxonomy across systems, so memories do not port.
• Combinatorial growth as new domains appear.
• Fails OKF’s interoperability goal — a generic consumer cannot interpret the types.

Risk Assessment:

• Technical Risk: Low to build, High to portability.
• Ecosystem Risk: High. Portability across systems is sacrificed.

Option 2: Flat tagging

Description: Drop typed categories entirely; let authors attach free-form tags to each memory.

Advantages:

• Zero upfront taxonomy design.
• Unlimited author flexibility.

Disadvantages:

• No semantic hierarchy: tags are keywords, not a classification.
• Poor organization; consumers cannot reason about a memory’s kind.
• Inconsistent vocabulary across authors degrades retrieval.

Risk Assessment:

• Technical Risk: Low.
• Ecosystem Risk: High. Without a shared kind-system, cross-system meaning collapses.

Option 3: Custom ontology per system

Description: Let every adopting system define its own ontology of concept types from scratch.

Advantages:

• Each system is internally coherent and fully tailored.

Disadvantages:

• No interoperability between systems — the opposite of MIF’s goal.
• High design burden duplicated by every adopter.
• No common base for tooling, decay defaults, or namespace conventions.

Risk Assessment:

• Technical Risk: Medium (each adopter re-solves the problem).
• Ecosystem Risk: High. Interoperability is lost.

Option 4: Cognitive-psychology memory triad (chosen)

Description: Adopt the cognitive triad from memory research as MIF’s base concept-type taxonomy:

1. Semantic — facts, concepts, relationships, and knowledge (declarative knowledge about the world: decisions, technology choices, preferences, domain knowledge).
2. Episodic — events, experiences, sessions, and timelines (time-bound records of “what happened”: incidents, conversations, blockers).
3. Procedural — step-by-step processes, runbooks, and patterns (how-to knowledge: migration guides, code patterns, workflows).

Domain-specific specialization is handled by ontology-extended types that declare a base of one of the three triad members (SPECIFICATION.md §4.2.1), keeping the base set fixed at three while permitting local vocabulary.

Technical Characteristics:

• Encoded as the conceptType enum ["semantic", "episodic", "procedural"] in the JSON Schema, with memoryType retained as a deprecated v0.1 alias.
• Each base type carries a namespace hint (semantic/*, episodic/*, procedural/*) that downstream conventions build on.
• Each typed concept remains a valid OKF concept: the type value satisfies OKF’s required field while supplying meaning OKF leaves open.

Advantages:

• Grounded in well-cited research — Tulving (1972) for the semantic/episodic distinction and Cohen & Squire (1980) for procedural memory as a distinct system.
• Universal: the three kinds map onto any domain’s facts, events, and processes.
• Intuitive mental model for both humans and AI systems.
• Enables meaningful cross-system memory portability under a single shared taxonomy.
• Directly answers OKF’s deliberately-absent concept-type question.

Disadvantages:

• Some memories span categories — mitigated by traits (see ADR-004).
• Assumes familiarity with the cognitive-psychology framing.
• Can feel constraining for narrow domain use cases — mitigated by ontology-extended types and namespace prefixes (see ADR-005).

Risk Assessment:

• Technical Risk: Low. Three-value enum is trivial to validate and stable.
• Schedule Risk: Low. The taxonomy is fixed and shipped.
• Ecosystem Risk: Low. A research-grounded, universal base maximizes portability and interoperability.

Decision

MIF adopts the cognitive triad — semantic, episodic, procedural — as its base concept-type taxonomy. These three base types are MIF’s opinionated answer to OKF’s deliberately-absent concept-type system.

Implementation specifics:

• The triad is the normative conceptType enum in the schema (schema/mif.schema.json), with memoryType kept as a deprecated v0.1 alias for backward compatibility.
• SPECIFICATION.md §4.2 defines the three base types and their namespace hints; §4.2.1 defines how ontologies extend them via a declared base.
• Specialization is achieved through ontology-extended types and underscore namespace prefixes (ADR-005), never by enlarging the base enum.

Consequences

Positive

1. Research-grounded foundation: The category boundaries rest on Tulving (1972) and Cohen & Squire (1980), giving reviewers an external authority to evaluate against.
2. Universal applicability: Facts, events, and processes exist in every domain, so the three base types apply everywhere.
3. Portability: A single shared taxonomy lets memories move between systems with preserved meaning.
4. OKF answer: MIF supplies the concept-type taxonomy OKF leaves open while each typed concept stays OKF-legible.

Negative

1. Cross-category memories: Some items legitimately span types; this is absorbed by traits rather than by adding base types (ADR-004).
2. Conceptual onboarding: Authors must understand the cognitive-psychology framing to type memories well.

Neutral

1. Fixed base, extensible edges: The base set is deliberately frozen at three; all domain growth happens through ontology-extended types and namespaces, by design (ADR-005).

Decision Outcome

The cognitive-triad taxonomy achieves the primary drivers: universality (facts / events / processes appear in every domain), theoretical grounding (Tulving; Cohen & Squire), and a concrete answer to OKF’s open concept-type question. Mitigations:

• Cross-category items are handled by the trait system (ADR-004) rather than by expanding the enum.
• Narrow domain needs are met by ontology-extended types and namespace prefixes (ADR-005), keeping the base set stable while allowing local specialization.

Related Decisions

• ADR-005: Underscore Namespace Prefix — the namespace prefix convention derives from the three base memory types.
• ADR-008: Decay Model Rationale — default decay half-lives vary by base memory type.
• ADR-004: Three-Tier Trait Inheritance — traits absorb memories that span more than one base type.
• ADR-009: OKF Compliance as a Superset — this base-type taxonomy is MIF’s answer to OKF’s deliberately-absent concept-type system.

Links

• Tulving, E. (1972). Episodic and semantic memory. In Organization of Memory — the semantic/episodic distinction.
• Cohen, N. J. & Squire, L. R. (1980). Preserved learning and retention of pattern-analyzing skill in amnesia: Dissociation of knowing how and knowing that. Science — procedural memory as a distinct system.
• Squire, L. R. (2004). Memory systems of the brain: A brief history and current perspective. Neurobiology of Learning and Memory.

More Information

• Date: 2026-06-18
• Source: SPECIFICATION.md “MIF answers OKF’s open questions” table and §4.2 (Memory Types); schema/mif.schema.json conceptType enum.
• Related ADRs: ADR-004, ADR-005, ADR-008, ADR-009, ADR-010
• Academic references: Tulving (1972); Cohen & Squire (1980); Squire (2004) — see Links above.

Audit

2026-06-18

Status: Compliant

Findings:

Finding	Files	Lines	Assessment
Base concept-type enum ["semantic", "episodic", "procedural"] is normative in the schema (conceptType)	schema/mif.schema.json	L30-L34	compliant
memoryType retained as a deprecated v0.1 alias of conceptType (same triad enum)	schema/mif.schema.json	L35-L38	compliant
Same triad enum enforced in the ontology schema	schema/ontology/ontology.schema.json	L122, L154	compliant
Three base memory types defined with descriptions and namespace hints	SPECIFICATION.md	L234-L250	compliant
Ontology-extended types specialize a declared base (keeps base set fixed at three)	SPECIFICATION.md	L252-L263	compliant
”MIF answers OKF’s open questions” table: triad is MIF’s answer to OKF’s absent concept-type taxonomy	SPECIFICATION.md	L49-L51	compliant

Summary: The cognitive triad is present and normative as the conceptType enum in schema/mif.schema.json, mirrored in the ontology schema, defined with namespace hints and extension rules in SPECIFICATION.md §4.2 / §4.2.1, and positioned in the SPECIFICATION’s “MIF answers OKF’s open questions” table as MIF’s answer to OKF’s deliberately-absent concept-type system.

Action Required: None.