holo-fiction

A memory where fiction is a first-class citizen.

Every item carries two orthogonal tags: a modal (the epistemic status — fact, hypothesis, fiction, counterfactual, or any user-registered category) and a world (the context the item belongs to, "real" by default). The substrate stores fiction and fact with the same primitives. The difference is the tag.

Three use cases motivated the design:

Knowledge-mode switch (DeepSeek-style). Distinguish "known science" from "frontier" from "speculative."
Book reader's memory. Track what is true in the world of a novel, cross-referenced against real-world facts.
Counterfactual reasoner. Ask "what would be the case if X were fact instead of fiction."

What it does

from holo_fiction import FictionalMemory

mem = FictionalMemory(d=2048)

# Real world
mem.observe("water_is_wet", modal="fact")
mem.observe("london", modal="fact", country="england")

# Fictional world
mem.observe("sherlock_holmes",
            modal="fiction", world="sherlock",
            occupation="detective", city="london")

# Separate queries
mem.query("water_is_wet", modal="fact")               # YES
mem.query("water_is_wet", world="sherlock")           # UNKNOWN
mem.query("sherlock_holmes", world="sherlock")        # YES

# Cross-world reference
mem.cross_world_reference("sherlock")
# [{"label": "sherlock_holmes", "field": "city",
#   "references": "london", "reference_world": "real"}]

# Character sheet
mem.character_sheet("sherlock_holmes", world="sherlock")

# What-if reasoning
mem.what_if("sherlock_holmes", "fiction", "fact",
            world="sherlock")

# Promotion
mem.promote("sherlock_holmes", "fiction", "fact",
            world="sherlock")

# Knowledge modes
mem.register_modal("known", confidence=1.0)
mem.register_modal("frontier", confidence=0.7)
mem.register_modal("speculative", confidence=0.4)

Results

All results at D=2048, threshold=0.05.

Self-test

Check	Result
bind/unbind identity	PASS
observe + query	PASS (conf=1.000)
fiction in separate world	PASS (conf=0.900)
fact world unaffected	PASS (conf=−0.014)

Fiction vs fact separation

Query	world=real	world=sherlock
water_is_wet	+1.015	−0.014
sherlock_lives_on_baker_street	+0.005	+0.907

Cross-contamination is at the noise floor.

Knowledge modes

Query	known	frontier	speculative
water_boils_at_100C	+1.016	−0.003	+0.006
universe_is_simulation	−0.022	−0.007	+0.411

Each query retrieves cleanly in exactly one mode.

Promotion

Storing dark_matter_exists as hypothesis (weight 0.5):

	fact	hypothesis	fiction
before promote	−0.035	+0.500	+0.000
after promote	+0.965	+0.000	+0.000

The item moves from one modal trace to another. No re-observation.

Character sheet

sherlock_holmes:
    occupation = detective  (fiction)
    friend     = john_watson (fiction)
    address    = 221b_baker_street (fiction)
    city       = london     (fiction)

Cross-world reference

Items in the fictional world that cite real-world labels:

sherlock_holmes     city -> london
john_watson         city -> london
221b_baker_street   city -> london

Installation

pip install numpy

No other dependencies. Single file, approximately 600 lines.

Usage

CLI

python holo_fiction.py
python holo_fiction.py --output results/

Runs ten demonstrations and writes a JSON state file.

Python

from holo_fiction import FictionalMemory

mem = FictionalMemory(d=2048)

# Storage
mem.observe(label, modal="fact", world="real",
            weight=1.0, source=None, **fields)

# Register a custom modal
mem.register_modal("known", confidence=1.0)
mem.register_modal("frontier", confidence=0.7)

# Query
result = mem.query(label, modal="fact", world="real",
                   fields=None)
# {"label": ..., "modal": ..., "world": ...,
#  "confidence": ..., "verdict": ...}

# Query across modals
results = mem.query_across_modals(label, world="real")

# Cross-modal operations
mem.what_if(label, from_modal, to_modal, world="real")
mem.promote(label, from_modal, to_modal, world="real")

# Analysis
mem.contradictions(world="real")
mem.modal_distribution(label)
mem.world_summary(world)
mem.character_sheet(name, world)
mem.cross_world_reference(world, reference_world="real")
mem.timeline(world, modal=None)

# Diagnostics
mem.stats()
mem.save_json("state.json")

API reference

`FictionalMemory`

FictionalMemory(d=2048, threshold=0.05)

Modal registration

register_modal(name, confidence=1.0) — add a modal with a confidence weight.
register_world(name) — add a world context.

Storage

observe(label, modal="fact", world="real", weight=1.0, source=None, **fields) — store an observation.

Query

query(label, modal=None, world="real", fields=None) -> dict
query_across_modals(label, world="real") -> {modal: result}

Cross-modal

what_if(label, from_modal, to_modal, world="real") -> dict
promote(label, from_modal, to_modal, world="real") -> dict

Analysis

contradictions(world="real") -> {modal_conflicts, field_conflicts}
modal_distribution(label, world=None) -> {modal: count}
world_summary(world) -> {counts, labels}
character_sheet(name, world) -> {fields: {field: [entries]}}
cross_world_reference(world, reference_world="real") -> [refs]
timeline(world, modal=None) -> [entries]

Diagnostics

stats() -> dict
save_json(path)

Design notes

Two orthogonal tags

Modal and world are independent. modal="fiction" in world="sherlock" is different from modal="fiction" in world="realm". The same item can be stored in multiple modals and multiple worlds; queries select by either tag.

Modal × world traces

Each (world, modal) pair has its own trace. The traces share the underlying codebook: name vectors and role vectors are global. This means an item stored in one (world, modal) has the same composite representation as the same item stored elsewhere. Promotion is a trace arithmetic operation, not a re-encoding.

Modal confidence as a weight

The confidence registered per modal multiplies the observation weight at storage time. A fiction item stored with weight 1.0 and modal confidence 0.9 produces a projection of ~0.9. The confidence factors are a design parameter, not a physical constant.

The field-matching principle

Observations with fields require queries with the same fields. If an item was stored with composite(a + b) and queried with composite(a), the projection dilutes.

The correct query:

mem.observe("dark_matter_exists", modal="frontier",
            evidence="rotation_curves")
mem.query("dark_matter_exists", modal="frontier",
          fields={"evidence": "rotation_curves"})

The demo's failure on dark_matter_exists is the empirical demonstration of this principle.

Cross-world reference

Field values share the name-vector namespace with labels. A field value that matches a label in the reference world is a cross-world reference. This works because both are stored as name vectors, not because of any special machinery.

Limitations

Field-matched queries required. Observations with fields must be queried with matching fields. Dilution, not error, is the failure mode.

Worlds are fully independent. No relation between worlds is modeled. A fiction world that is "about" the real world is not distinguished from one that is not.

Modal confidence fixed at registration. Changing a modal's confidence does not retroactively re-weight items. Re-storage is required.

No modal inference. Every observation specifies its modal explicitly. The substrate does not infer the modal from context.

No cross-world causal structure. A fictional world cannot declare that its events have effects in the real world, or vice versa.

promote mutates records. The Obs.modal field is mutated in place. Copy the record if audit history is required.

Citation

@misc{holo-fiction2026,
  title  = {holo-fiction: A memory where fiction is a first-class
            citizen},
  author = {zeechimp},
  year   = {2026},
  note   = {Modal and world as orthogonal tags on every item.
            Knowledge modes, book reader's memory, and
            counterfactual reasoning.}
}

References

Plate, T. A. "Holographic Reduced Representations." IEEE Transactions on Neural Networks 6:3 (1995), 623–641.
Kanerva, P. "Hyperdimensional Computing." Cognitive Computation 1:2 (2009), 139–159.
Lewis, D. "Truth in Fiction." American Philosophical Quarterly 15:1 (1978), 37–46.
Ryan, M.-L. "Possible Worlds, Artificial Intelligence, and Narrative Theory." Indiana University Press (1991).

License

Apache 2.0

Downloads last month: -; Downloads are not tracked for this model. How to track