A Story World Must Not Forget What the Model Has Not Mentioned

An AI assistant and a persistent story world operate under two different kinds of information physics. For an assistant, an old preference may deserve less attention than a recent one. For a story world, a fact does not become less true because hundreds of scenes have passed without mentioning it.

If Building A has a north wing, the wing should not slowly disappear from the system because retrieval stopped surfacing it. It exists until something in the world changes that state.

The architectural rule is simple:

Retrieval relevance may decay. Ontological existence does not.

That distinction is central to any system intended to keep a fictional universe coherent over long timelines, revisions and multiple media.

Why Decay Makes Sense for Memory

Memory and recommendation systems need selection. A model cannot place every previous observation into every new context, so the system ranks what it recalls. Recency is often a useful signal:

retrieval_score(x, query, now)
= semantic_similarity
+ importance
+ entity_match
+ recency_weight

A preference stated yesterday may be more useful than one recorded five years ago. A recently successful workflow may be more relevant than an old attempt made under different conditions. In these cases, decay helps the system allocate limited attention.

But a lower retrieval score means only:

this information is less useful for the current query

It must not silently mean:

this information is becoming less true

Confusing those two statements turns a context-management technique into an accidental model of reality.

Recommendation Needs Two Timescales

A recommendation system that must feel both immediately relevant and recognisably personal cannot rely on one undifferentiated history. It needs at least two logical forms of memory:

user model
= long-term attractors
+ short-term activation

The long-term layer represents interests that have been reinforced across time: recurring subjects, communities, formats or creators. The short-term layer represents what has become salient during the current session, day or brief phase of attention.

Either layer alone produces a caricature of the user. Long-term history without decay can keep serving a three-month interest from 2017 as if it still defined the person in 2030. Short-term activation without a stable baseline can interpret five reactions to one television series as a complete change of identity.

The combined behaviour is more useful:

not reinforced for a long time → weight decreases
repeatedly reinforced          → weight increases
sudden interaction burst       → temporary feed boost

The implementation might involve embeddings, graph features, candidate retrieval, ranking models and many additional signals. Yet the observable result can remain surprisingly simple: show more of what the person is engaging with now, embedded within what they tend to engage with over time.

This also explains why a feed may appear to have a historical horizon without containing a literal rule such as “forget everything before 2021.” Sufficient temporal decay, combined with a much larger volume of recent interactions, can produce the same emergent effect:

weight(old interest) ≈ 0

The interest is not necessarily deleted. It has become inactive and may return if new interactions reinforce it. A radically newer recommendation architecture can therefore reproduce the same macroscopic dynamics as an older feed. That is not automatically regression; it may be a stable attractor of personalised recommendation itself.

This is precisely where the analogy must end. An inactive preference and a nonexistent building are not equivalent states.

World State Persists Until an Event Changes It

Suppose the story establishes:

Building A HAS_PART North Wing

After 400 sequences without a reference to that location, the relationship is still valid. The system may omit it when planning a conversation in another city. It must retrieve it when a character returns to Building A, when a floor plan is generated, or when a continuity check asks what can be seen from the courtyard.

The truth condition is therefore not exponential decay. It is an interval:

HAS_PART(Building A, North Wing)
valid_from: sequence 12
valid_to: open

The open end means “valid until explicitly changed,” not “probably valid until the model forgets.” This corresponds to time-based graph versioning, where entities and relationships receive defined validity intervals and historical graph states remain queryable.

Destruction Creates a New State, Not a Weaker Fact

Now imagine that the north wing is destroyed in Sequence 417. The earlier statement should not be down-weighted until it vanishes. It remains historically true, while a new state becomes true after the event:

Sequence 12–416:  North Wing exists
Sequence 417:     North Wing destroyed
Sequence 418–…:   North Wing no longer exists

This lets the system answer two different questions correctly:

• “Can the characters enter the north wing now?” — No.
• “Where did Mara meet Elias in Sequence 203?” — In the north wing, which existed then.

An overwrite-only database may preserve the first answer and lose the second. A similarity-only memory system may retrieve either statement without understanding which time it governs. A temporal world model preserves identity, event and validity together.

Event sourcing offers a useful complementary pattern: record state changes as events so earlier states can be reconstructed. A story system does not need to implement every part of formal event sourcing to learn from the principle. “North wing destroyed” should be an accepted event with an actor, a narrative time, provenance and explicit consequences—not merely a new paragraph that competes with an old paragraph in a vector search.

Two Axes, Not One Score

The system needs at least two independent judgments:

Validity should filter candidate facts before relevance ranks them. Otherwise, a highly similar but superseded fact may outrank the current truth, while an old but still valid fact may be excluded merely because it has not appeared recently.

This is why memory is not state . Memory helps assemble a useful model context. State constrains what the model is allowed to treat as true.

A Practical Retrieval Pipeline for Story Worlds

A continuity-aware system can assemble context in a deliberate order:

request + requested story time
→ resolve stable entities
→ select facts valid at that time
→ traverse causally and narratively relevant relations
→ rank the valid results for the current task
→ fit the selected facts into the model context

This design still permits forgetting where forgetting is healthy. Temporary observations can expire. Unapproved ideas can be discarded. Personal preferences can be revised or deleted. Embeddings can be regenerated. Summaries can be replaced.

Canon requires a different operation: an authorized transition. A fact may be corrected, superseded, scoped to one version of the story, or ended by an event. Each change should remain traceable.

What This Means for Saganode

Saganode describes a connected story world in which characters, places, objects, lore and storylines remain available as the world evolves. That promise depends on preserving the difference between absence from the current context and absence from the world.

The ontology or world-state layer should therefore own:

• stable identities for characters, locations, objects and concepts;
• relationships with explicit temporal validity;
• accepted events that create, alter or end those relationships;
• historical states that remain queryable;
• branch or canon scope when alternative versions coexist; and
• provenance for who or what approved a change.

The retrieval layer can then remain aggressively selective. It may prefer the current location, active characters, unresolved conflicts and facts connected to the scene objective. It can ignore ten thousand unrelated truths without deleting any of them.

This separation also clarifies the role of the generative model. The model may propose that the north wing is destroyed. It may draft the scene, infer consequences or generate visuals. The change enters the world only when the production system accepts an explicit transition. The broader studio architecture owns the world; the model explores possible changes to it.

The World Does Not Forget. It Changes.

A coherent story universe needs both selective attention and durable truth. Without selective retrieval, context becomes unusably large. Without ontological persistence, the universe becomes whatever the model happened to mention most recently.

The right contract is:

old information in memory → may receive less retrieval weight
old valid fact in the ontology → remains true
accepted world event → closes one validity interval and opens another

That is the difference between remembering a story and maintaining a world.