VSABrains

VSA vs VSABrains

VSABrains is not “anti-VSA.” It is a response to a specific failure mode we observed when using holographic superposition to represent long narratives: order becomes muddy. This page explains the tradeoff plainly and shows where VSA ideas still exist in the system.

The core difference: superposition vs address

Classic Vector Symbolic Architectures (VSA) represent many items in a single vector by superposing them. This is elegant and algebraic, but under long sequences it tends to blur order. VSABrains chooses a different encoding: order as address. Steps are written along a trajectory in a discrete map, so “when” is stored as “where.”

Comparison diagram: VSA superposition versus VSABrains discrete trajectories

Where do frames fit?

Frames are a semantic layer above both systems. In a classic VSA, you might bind “emotion” and “theme” into vectors. In VSABrains, frames are explicit counters and relations attached to events. They are easier to audit and explain to a non-ML audience.

Frames do not replace trajectories. They answer “what is dominant?” while trajectories answer “where was it written?” See Frames & CNL for the plain-language guide.

What VSABrains keeps from VSA

  • Symbolic tokens: events become stable token IDs.
  • Binding by composition: a step token is a deterministic hash of role tokens.
  • Distributed evidence: heavy-hitters retain multiple candidates per location.

In short: we still compose symbols, but we do not rely on dense holographic superposition to carry timeline structure.

Why the discrete choice helps

Auditable order
A step has a concrete location. Retrieval is a replay plus an index lookup, not a fragile unbinding guess.
Consensus under noise
Multiple columns can disagree and then vote. A single holographic stream has no native multi-view consensus.
Index beats scan
Localization uses a token→location index. A naive list must scan linearly over the full history.
Graceful degradation
Heavy-hitters cap local memory and make loss explicit. Superposition tends to degrade invisibly.

What this does not claim

  • It does not claim holographic memory or perfect unbinding from a single endpoint.
  • It does not claim lower ingestion cost than a simple list (it often costs more).
  • It does claim that, at scale, indexed localization + checkpoints can reduce query cost.

See Experiment 4 for consensus under noise, and Experiment 5 for scaling behavior on larger histories.

One-sentence summary

VSA compresses by superposing content into a shared space; VSABrains compresses by indexing trajectories in addressable space. The first is algebraically beautiful; the second is operationally auditable.