Why VSABrains?

VSABrains is built to make temporal order auditable and robust under noise. It encodes order as spatial address (trajectories in discrete maps), then uses multi-column consensus to reduce ambiguity. This page answers common questions and clarifies what the architecture does and does not claim.

Is this still VSA? Is there holography?

It is inspired by VSA ideas (symbol composition, binding, and distributed representations) but it is not a holographic VSA system. VSABrains does not rely on dense superposition to represent long sequences. Instead, it uses discrete maps and heavy-hitters to keep top-K tokens per cell, which is sparse and auditable.

So: no holography claim. The architecture is a different tradeoff: discrete, CPU-friendly, and traceable, even if it sacrifices some of the smooth algebra of classic VSA.

What is a column, in simple terms?

A column is a small independent witness. It walks across a grid and writes tokens into cells. Multiple columns see the same story but from slightly different coordinate systems.

• Why multiple? So they can disagree and then vote (consensus).
• Why grids? So time becomes location and can be indexed.

What are frames?

Frames are semantic lenses that sit on top of events. Think of them as extra tags like “emotion,” “theme,” “dialogue act,” or “conflict type.” They do not replace columns; they help answer higher-level questions.

Frames are defined by a CNL profile (a small, constrained list) and then updated at each step. See Frames & CNL for the full explanation.

Is the meaning of a story its final point?

No. The final location is not the “meaning.” It is just the end of a trajectory. The point of VSABrains is that the path is written into memory as a sequence of locations. Meaning comes from replay and evidence chains, not from a single endpoint.

Can we reconstruct the path from the final point?

Not from the final point alone. A single endpoint does not encode the full path. Path reconstruction is done through:

• Replay from checkpoints and the episodic store.
• Localization using recent token windows (fast maps + location index).
• Slow maps as coarse summaries for long horizons (approximate).

This is a deliberate design: store enough structure to replay deterministically, and avoid pretending that a final point contains all history.

Can we compress columns and still reconstruct paths?

Yes, but only approximately. The system already supports compression via checkpointing (store partial state) and slow-map summaries (store coarse tokens for older history). This enables approximate reconstruction without retaining every detail.

If perfect replay is required, you must keep the full episodic stream. If approximate replay is acceptable, checkpoints + slow maps can reduce storage at the cost of fidelity.

Why not just keep a list of events?

A pure list is strong for small data, but it has tradeoffs:

• Scaling: retrieval is linear unless you build separate indexes.
• Noise: a single stream offers no multi-view consensus.
• Ambiguity: a list has no native mechanism to maintain parallel hypotheses.
• Auditability: you still need explicit checks for conflicts, otherwise you return inconsistent answers without warning.

VSABrains makes these behaviors explicit: spatial indexing for fast localization, multi-column consensus for robustness, and replay + verifiers for auditable answers.

What questions can it answer well?

• Where someone is (now or at a step)
• Who has an item
• Whether a contradiction occurred
• Dominant emotion, theme, or tone (if frames are enabled)

These are strong because they map to indexed locations and explicit frame counters.

What questions are hard?

• Deep causal reasoning without explicit rules
• Unstated implications or “mind-reading”
• Questions that require facts not in the story

Where is the evidence?

See Experiment 4 in eval/exp4-consensus. It compares a single noisy stream and a naive list-scan baseline against multi-column consensus under the same noise rate. Consensus should yield higher accuracy and lower disagreement, while the list baseline shows the cost of linear scanning.