Skip to content
MIF

Decay Model Rationale (Configurable Temporal Freshness)

ADR-008: Decay Model Rationale (Configurable Temporal Freshness)

Section titled “ADR-008: Decay Model Rationale (Configurable Temporal Freshness)”

Status

Section titled “Status”

Accepted

Context

Section titled “Context”

Background and Problem Statement

Section titled “Background and Problem Statement”

A modeled-information bundle accumulates knowledge whose currency erodes at different rates. Consumers of a MIF bundle need mechanisms to:

  • Prioritize recent and relevant knowledge over aged knowledge.
  • Express that some facts go stale quickly while others remain durable.
  • Bound storage growth so that low-value, decayed entries can be reclaimed.
  • Distinguish freshness (how current a fact is) from validity (the interval in which a fact is asserted to hold).

The exponential-decay curve has an established analogue in the cognitive-science account of human memory, where retained items lose strength over time unless reinforced. That analogue supplies an intuitive, well-understood mental model for relevance ranking, but the core MIF model frames the function purely as freshness — the cognitive-memory framing now lives in the AI Memory profile.

Current Limitations

Section titled “Current Limitations”
  • OKF defines a directory of .md concept files but deliberately leaves stale-vs-live handling to the consuming process — it offers no freshness or validity model at all.
  • Without a content-model answer, “is this knowledge still current?” is a question every consumer must re-invent, with no interoperable contract.
  • A single fixed decay rate cannot serve both fast-moving episodic records and slow-moving semantic facts; the model must be configurable per entry and per kind.

Decision Drivers

Section titled “Decision Drivers”

Primary Decision Drivers

Section titled “Primary Decision Drivers”
  1. Freshness semantics: The temporal model must express how current a piece of knowledge is, independent of whether it is still valid.
  2. Configurability: Decay behavior must be selectable per entry (model + half-life) so different knowledge kinds decay appropriately.
  3. OKF gap-filling: Decay/validity is MIF’s opinionated answer to OKF’s open stale-vs-live question; it must be a concrete, documented contract.

Secondary Decision Drivers

Section titled “Secondary Decision Drivers”
  1. Storage management: Decayed, low-strength entries should become candidates for archival, compression, or deletion.
  2. Reinforcement: Access or explicit reinforcement should be able to reset or slow decay, mirroring a familiar mental model.
  3. Standards alignment: Durations use ISO-8601 so half-lives and TTLs are unambiguous and portable.

Considered Options

Section titled “Considered Options”

Option 1: No temporal model (defer stale-vs-live to the consumer, as OKF does)

Section titled “Option 1: No temporal model (defer stale-vs-live to the consumer, as OKF does)”

Description: Carry no decay or freshness data; leave every consumer to decide whether a fact is still current.

Advantages:

  • Minimal surface; nothing to specify or validate.
  • Matches OKF’s bare interoperability surface exactly.

Disadvantages:

  • Re-creates the OKF gap MIF exists to fill — no interoperable freshness contract.
  • Relevance ranking and garbage collection have no shared basis.

Risk Assessment:

  • Technical Risk: Low to build, but high to the value proposition — the content model is the reason to adopt MIF over bare OKF.
  • Ecosystem Risk: High. Every consumer diverges on staleness handling.

Option 2: A single fixed decay rate

Section titled “Option 2: A single fixed decay rate”

Description: Apply one global exponential decay with a fixed half-life to all entries.

Advantages:

  • Simple to implement and reason about.

Disadvantages:

  • One rate cannot fit both fast-moving events and durable facts.
  • No way to mark a fact permanent or to model hard expiration.

Risk Assessment:

  • Technical Risk: Low.
  • Ecosystem Risk: Medium. Forces a single tuning onto every knowledge kind.

Option 3: A configurable decay model (chosen)

Section titled “Option 3: A configurable decay model (chosen)”

Description: A per-entry decay object selects one of four models (none / linear / exponential / step) over an ISO-8601 halfLife, with a currentStrength in [0,1]. Default half-lives differ by knowledge kind, access can reinforce an entry, and weak entries become reclamation candidates.

Technical Characteristics:

  • Decay object lives under temporal, alongside validFrom / validUntil / recordedAt / ttl, separating freshness from the validity window.
  • The freshness curve strength = e^(-t/halfLife) is fully current at record time and decays gradually toward stale.

Advantages:

  • Per-entry and per-kind tuning; none supports permanent knowledge.
  • Directly answers OKF’s open stale-vs-live question with a concrete contract.
  • Reinforcement and garbage-collection behaviors fall out naturally.

Disadvantages:

  • Adds management complexity (background decay calculation, tuning).
  • Mis-tuned half-lives can deprecate still-important knowledge.

Risk Assessment:

  • Technical Risk: Low. The fields are bounded and schema-validated.
  • Schedule Risk: Low. The model shipped in v1.0.0.
  • Ecosystem Risk: Low. A documented, machine-checkable freshness contract.

Decision

Section titled “Decision”

Implement a configurable decay model under the temporal object.

Decay Types

Section titled “Decay Types”
decay:
  model: "exponential"  # none | linear | exponential | step
  halfLife: "P14D"      # ISO 8601 duration
  currentStrength: 0.85 # 0.0-1.0

Supported Models

Section titled “Supported Models”
  1. none — Knowledge never decays (permanent).
  2. linear — Strength decreases at a constant rate.
  3. exponential — Strength follows an exponential decay curve.
  4. step — Strength drops at a defined threshold (hard expiration).

Default Half-Lives by Knowledge Kind

Section titled “Default Half-Lives by Knowledge Kind”
Knowledge KindDefault Half-LifeRationale
SemanticP30D (30 days)Facts change slowly
EpisodicP7D (7 days)Events become less relevant
ProceduralP14D (14 days)Processes need periodic refresh

Reinforcement

Section titled “Reinforcement”

Accessing or explicitly reinforcing an entry resets decay:

temporal:
  last_accessed: "2026-01-27T10:00:00Z"
  access_count: 5
  last_reinforced: "2026-01-25T15:00:00Z"

Garbage Collection Threshold

Section titled “Garbage Collection Threshold”

Entries with currentStrength < 0.1 are candidates for archival (move to cold storage), compression (reduce to summary), or deletion (with user consent).

Relationship to OKF

Section titled “Relationship to OKF”

Decay/validity is MIF’s opinionated answer to OKF’s deliberately-open stale-vs-live question. In the “MIF answers OKF’s open questions” table, “Stale-vs-live left to process” is answered by “Validity windows + TTL/freshness.” The decay function expresses freshness; the validFrom/validUntil window bounds the interval in which a fact is asserted to hold. This is the same superset posture ADR-009 establishes: MIF supplies the content model OKF leaves open.

Consequences

Section titled “Consequences”

Positive

Section titled “Positive”
  • Automatic relevance ranking based on recency.
  • Storage management via garbage collection of weak entries.
  • Mirrors a familiar mental model (decay-and-reinforce).
  • Configurable per entry and per kind.
  • Supports permanent knowledge via model: none.
  • Provides the interoperable freshness contract OKF lacks.

Negative

Section titled “Negative”
  • Adds complexity to bundle management.
  • Requires background decay calculation.
  • May accidentally deprecate important knowledge if mis-tuned.
  • Tuning half-lives requires experimentation.

Neutral

Section titled “Neutral”
  • Freshness (decay) and validity (validFrom/validUntil) are separate concerns carried in the same temporal object — by design.
  • The cognitive-memory origin of the exponential curve is retained as rationale but moved to the AI Memory profile, keeping the core framed as freshness.

Decision Outcome

Section titled “Decision Outcome”

The configurable-decay approach achieves the primary drivers: freshness semantics (the e^(-t/halfLife) curve), configurability (four models over a per-entry half-life), and OKF gap-filling (a concrete stale-vs-live contract). Mitigations:

  • Default half-lives are documented as pragmatic, non-prescriptive defaults that implementations SHOULD tune by knowledge kind, organizational velocity, and access patterns.
  • Reinforcement (lastAccessed / accessCount) lets implementations slow or reset decay so frequently-used knowledge is not prematurely deprecated.

Implementation Notes

Section titled “Implementation Notes”

Decay Calculation

Section titled “Decay Calculation”
currentStrength = initialStrength * (0.5 ^ (elapsed / halfLife))

Note: The canonical half-life formula is 0.5^(t/halfLife). The specification’s e^(-t/halfLife) is the natural exponential decay form. For true half-life behavior (strength = 0.5 when t = halfLife), use: e^(-ln(2) * t / halfLife) which equals 0.5^(t/halfLife).

Garbage Collection Threshold

Section titled “Garbage Collection Threshold”

Entries with currentStrength < 0.1 are candidates for:

  • Archival (move to cold storage)
  • Compression (reduce to summary)
  • Deletion (with user consent)

Reinforcement Triggers

Section titled “Reinforcement Triggers”
  • Direct access (read)
  • Explicit reinforcement command
  • Citation by other entries
  • User interaction referencing the entry
Section titled “Related Decisions”
Section titled “Links”

More Information

Section titled “More Information”
  • Date: 2026-01-27
  • Source: SPECIFICATION.md §9 (Temporal Model) and the “MIF answers OKF’s open questions” table; schema/mif.schema.json (temporal.decay).
  • Related ADRs: ADR-001, ADR-009

Audit

Section titled “Audit”

2026-06-18

Section titled “2026-06-18”

Status: Compliant

Findings:

FindingFilesLinesAssessment
Temporal Model section exists; decay is a temporal property alongside validFrom/validUntil/recordedAt/ttlSPECIFICATION.mdL1117-L1134compliant
Four decay models enumerated (none/linear/exponential/step) with exponential = e^(-t/halfLife)SPECIFICATION.mdL1136-L1143compliant
Freshness rationale: decay answers OKF’s open “live vs. stale” question; defaults P7D/P14D/P30D are non-prescriptive; reinforcement via lastAccessed/accessCountSPECIFICATION.mdL1145-L1172compliant
Cognitive-memory origin moved to AI Memory profile; core framed as freshnessSPECIFICATION.mdL1174-L1179compliant
”MIF answers OKF’s open questions” table: “Stale-vs-live left to process” → “Validity windows + TTL/freshness”SPECIFICATION.mdL49-L55compliant
Abstract states MIF supplies “validity/freshness semantics” as its OKF content-model fillSPECIFICATION.mdL18-L22compliant
Schema temporal.decay: model enum [none, linear, exponential, step], halfLife (pattern ^P), currentStrength (0-1), lastReinforcedschema/mif.schema.jsonL297-L325compliant
JSON-LD context terms decay, model, halfLife (xsd:duration), currentStrength (xsd:decimal), lastReinforcedschema/context.jsonldL158-L169compliant
GC threshold currentStrength < 0.1 is an ADR-level decision; spec defines a related compression trigger at Strength < 0.3 AND content > 100 lines, not a 0.1 GC thresholdSPECIFICATION.mdL604-L605partial

Summary: The configurable decay model, its four models, the ISO-8601 half-life, the per-kind defaults, reinforcement, and the OKF stale-vs-live linkage are all present and normative in SPECIFICATION.md, mif.schema.json, and context.jsonld. The currentStrength < 0.1 garbage-collection threshold is an ADR-level operational decision and is not stated verbatim in the specification — the closest spec text is the Strength < 0.3 compression trigger (a different threshold for a different action), so that one finding is marked partial.

Action Required: None for the decision itself. If the < 0.1 GC threshold is intended to be normative, it should be added to SPECIFICATION.md §5.6 alongside the existing compression criteria.