Memory tier classification

[jimador]

Observation

All propositions in a context are treated identically for retrieval, decay, and eviction. In practice, propositions have different temporal relevance:

• Recently accessed or reinforced propositions are actively relevant
• Propositions from earlier in the conversation but still occasionally referenced are warm
• Propositions that haven't been accessed in a while but aren't contradicted are cold

Without tier classification, decay and eviction treat all propositions the same:

• Too aggressive → active propositions fade too quickly
• Too conservative → stale propositions consume token budget

What DICE already has

• Proposition.lastAccessed: Instant — tracked but not used for behavior differentiation
• Proposition.reinforceCount: Int — frequency signal, not used for tier classification
• Proposition.decay: ZeroToOne — uniform rate, not modulated by access patterns
• MemoryMaintenanceOrchestrator — runs maintenance but doesn't classify propositions by recency
• PropositionRepository — has findByRevisedBetween() and findAllOrderedByEffectiveConfidence(), but no tier-based queries

The question

Should propositions have a temporal tier that modulates their decay and eviction behavior?

A simple model:

Tier	Criteria	Effect
HOT	Accessed or reinforced recently	Slower decay, higher eviction resistance
WARM	Not recently accessed, but not stale	Normal decay
COLD	No access in a while	Faster decay, first to evict

Implementation options

1. Computed on-the-fly — classify based on lastAccessed relative to configurable thresholds. No storage, always fresh. Cost: recalculated on every query.

2. Stored field — add memoryTier: MemoryTier to Proposition. Updated during maintenance sweeps. Enables tier-based repository queries, but can go stale between sweeps.

3. Decay multiplier only — no explicit tier. lastAccessed directly modulates the decay rate in effectiveConfidence(). Simpler, but loses the categorical abstraction.

Open questions

• Is lastAccessed sufficient for classification? Other signals matter: reinforceCount recency, whether the proposition is referenced by active abstractions, whether the conversation thread it came from is still ongoing.
• Should tier be stored or computed? Stored enables efficient queries (findByTier(HOT)) but risks staleness. Computed is always fresh but can't be indexed.
• Is this redundant with decay? If effectiveConfidence() already decays based on revised time, adding a tier multiplier based on lastAccessed creates two time-based decay dimensions. Is that useful or confusing?

[igordayen]

@jimador - speaking about temporal representation, it could be modeled after spring framework:

Auditable<U, ID, T extends [TemporalAccessor]

with extension of accessedBy/lastAccessedDate.

Entities / Propositions can be sorrted by date, classification can use "by-date"segments.

Perhaps decay as function takes lastAccessedDate as parameter.

More complete temporal memory representation can track versioned (upon modification) Entities too.

There is also close similarity between proposed tiers and memory types as LTM (long-term memory) and STM (short-term memory).

In summary, IMO, lastAccessedTimastamp is primary value, while decay gets continiousely computed. In the very end, main objective is to compile candidates for working memory to be injected into prompt dynamically.
Still learning, is "attention" as a concept - yet another sub-domain of a working memory?