← Back to Live Labs

ARTI / STATEFUL RECALL / REAL WASM

Remember one signal. Recognize its near match. Reset without changing weights.

Run a stateless baseline, commit the reference trace to fixed-capacity session state, then challenge it. The real read/update graphs show a near match becoming familiar while a novel signal stays novel.

Parametersunchanged→Session state4 slots · 144 B→Any timefork / reset

01 / THE EVIDENCE

Three visible signals. One session memory.

Each card is the actual [1,2,4] trace received by the model. No label or hidden text enters inference.

Not started
Reference signalcommitted to the session
2.00 · 0.00 · 0.00 · 0.00 · 1.80 · 0.00 · 0.00 · 0.00
Near matchsmall perturbation
2.00 · 0.03 · 0.00 · 0.00 · 1.80 · -0.02 · 0.00 · 0.00
Novel signaldifferent structure
2.00 · 0.80 · 0.00 · 0.00 · 1.80 · 0.72 · 0.00 · 0.00

02 / RUN THE EXPERIMENT

Establish the baseline, then commit memory

  1. 1Start the session and read all three signals with empty state.
  2. 2Atomically commit the reference trace to fixed-capacity state.
  3. 3Probe the near and novel signals again and compare live responses.

03 / WHAT CHANGED

Same model. Different session state.

Before memory · Near match—Session match response
After memory · Near match—Session match response
After memory · Novel signal—Session match response
○
Waiting for a real run

If execution fails, no precomputed result appears here.

Engineering evidence · real state slots, values, and runtime boundary

EXPLICIT STATE

4 fixed-capacity state slots

S00.000
S10.000
S20.000
S30.000

LIVE READOUT

Current probe

Session match response
—
Recall contribution
—
STATE
0 B
PROVIDER
—

Inputs come from the public parity fixture; read.onnx and update.onnx execute live in this browser. Match response is this Recall graph's retrieval gate—not general accuracy or semantic probability.