Temporal Model

MIF uses a bi-temporal model distinguishing between:

Transaction Time: When the memory was recorded in the system
Valid Time: When the fact represented by the memory is true

Temporal Properties

Property	Type	Description
`validFrom`	DateTime	When fact becomes valid
`validUntil`	DateTime	When fact expires (null = indefinite)
`recordedAt`	DateTime	When recorded (transaction time)
`ttl`	Duration	Time-to-live (ISO 8601 duration)
`decay`	Object	Decay model parameters
`accessCount`	Integer	Times accessed
`lastAccessed`	DateTime	Last access time

Decay Models

Model	Formula	Use Case
`none`	No decay	Permanent memories
`linear`	strength = 1 - (t / ttl)	Simple linear decay
`exponential`	strength = e^(-t/halfLife)	Natural forgetting curve
`step`	strength = 1 if t < ttl else 0	Hard expiration

Decay Rationale

MIF’s decay model values (P7D, P14D, P30D half-lives) are pragmatic defaults for AI memory systems, inspired by but not directly derived from cognitive psychology research. They represent reasonable approximations for memory management in agentic contexts.

Scientific Background

The exponential decay model strength = e^(-t/halfLife) is inspired by Hermann Ebbinghaus’s forgetting curve (1885), which demonstrates that memory retention follows an exponential decline:

Time Elapsed	Approximate Retention
1 hour	~50%
24 hours	~30-35%
7 days	~25%
30 days	~10%

The mathematical form R = e^(-t/S) where R is retrievability, t is time elapsed, and S is memory strength, has been validated by modern replication studies.

Why These Specific Values?

Half-Life	Use Case	Rationale
P7D	Short-term context	Aligns with weekly work cycles and episodic memory consolidation windows
P14D	Medium-term projects	Spans typical sprint/iteration boundaries
P30D	Long-term knowledge	Corresponds to monthly review cycles and hippocampal consolidation periods (~30 days in animal studies)
P90D	Default TTL	Quarterly relevance for most organizational knowledge

These values are not prescriptive—implementations SHOULD tune them based on:

Memory type (episodic decays faster than semantic)
Organizational context (high-velocity vs. stable environments)
Access patterns (frequently accessed memories reinforce slower decay)

Memory Consolidation Considerations

Research on memory consolidation suggests memories transition from hippocampus-dependent (recent) to cortex-dependent (remote) storage over time. MIF’s lastAccessed and accessCount fields enable implementations to model reinforcement—each access can reset or slow decay, analogous to spaced repetition strengthening memory traces.

References:

Ebbinghaus, H. (1885). Memory: A Contribution to Experimental Psychology
Murre & Dros (2015). Replication and Analysis of Ebbinghaus’ Forgetting Curve
Squire & Bayley (2007). The neuroscience of remote memory
Wickelgren (1972). Trace resistance and the decay of long-term memory

Example

temporal:
  valid_from: 2026-01-15T00:00:00Z
  valid_until: null
  recorded_at: 2026-01-15T10:30:00Z
  ttl: P90D
  decay:
    model: exponential
    halfLife: P7D
    strength: 0.85
    last_reinforced: 2026-01-18T09:00:00Z
  access_count: 5
  last_accessed: 2026-01-20T14:22:00Z