Dynamic Size Representations (Research Theme)

Home · Architecture · Theory · Specs · Research · Dynamic Size Why “elastic” representations matter for capacity, determinism, and long-running sessions
Definition (in this project): a “dynamic size” (or “elastic geometry”) HDC representation is one where the effective vector width can grow as new atoms appear, without forcing a global re-encode of previously loaded theories. The critical property is prefix stability: earlier encodings remain valid when the representation expands.

Why this is important

Many VSA/HDC systems choose a fixed dimension (e.g. 2048–16384 bits) and accept that large knowledge bases eventually reduce retrievability due to superposition noise. In AGISystem2 we want to study what happens when we remove (or relax) that fixed geometry constraint and allow the representation to grow with the knowledge base.

Capacity under growth

When KB size increases, “one fixed dimension” forces a trade-off between speed and collision/noise. Elastic size lets us explore a different trade-off: pay additional bytes/bits only when the universe expands.

Determinism across runs

Evaluators often run many sessions sequentially in the same Node process. A session-local allocator plus elastic geometry keeps results reproducible without relying on a process-global dictionary.

Reasoning ergonomics

“Exact-ish” witnesses (role markers, position atoms, operator families) become easier to detect when the representation can preserve structure and avoid destructive compression.

Strategies in AGISystem2 that support dynamic size

Strategy What grows Why it helps Spec
Metric-Affine Elastic (EMA) Byte geometry grows by chunks Superposition stays usable under larger KBs; bundling can be chunk-aware DS23
EXACT BigInt bitsets naturally extend to higher bit indices Lossless membership / witness extraction; upper bound for retrievability DS25
Sparse-Polynomial Exponent set space is not tied to a fixed bit-width Compact storage at low k; useful baseline for “growth without fixed dimension” DS15
Dense-Binary (baseline) Does not grow (fixed bits) Reference point for classic HRR/VSA behavior DS09
Metric-Affine (fixed) Typically fixed bytes per vector Fast baseline for byte-channel geometry before adding elasticity DS18
Important: “dynamic size” does not automatically mean “more accurate”. It changes the cost model (time/memory) and the failure modes. A major research question is when elastic growth is worth it, and when a fixed geometry + better cleanup/decoding is the better engineering choice.

Biological intuition (careful analogy)

We treat this as an analogy rather than a strict claim about neurobiology: biological systems do not appear to commit to a single global “vector width” that is fixed forever. Instead, coding capacity can be supported by:

In AGISystem2, EMA and EXACT are the “engineering analogs” of this idea: expand capacity as the KB grows, rather than fixing a dimension upfront and accepting progressive interference.

How we evaluate it

The main experimental harnesses are:

Open questions