braintrace

Build the model with braintrace.nn ETP-aware layers, compile once with braintrace.compile(model, braintrace.D_RTRL, example_step), then call learner.etrace_grad() to accumulate online gradients. Memory stays constant in sequence length because eligibility traces replace the unrolled graph.

Use D_RTRL when parameter-shaped trace memory O(B * |theta|) is feasible; it is the general RNN default. Use pp_prop when parameter-shaped traces are too large, especially for recurrent SNNs, since it stores input/output factors at O(B * (I + O)).

Not in general. D_RTRL uses a diagonal hidden-Jacobian approximation and pp_prop adds input/output factorization, so equality with BPTT holds only in documented mathematical regimes. Validate against a reduced BPTT oracle before relying on gradient fidelity.

A parameter only participates when consumed by an ETP operator; ordinary JAX operations exclude it automatically. Readouts that do not feed hidden State are correctly reported as non-temporal, and weights reaching hidden State only through another trainable ETP weight are excluded to prevent double counting.

Call both brainstate.nn.reset_all_states(model, batch_size=...) and learner.reset_state(batch_size=...) at every independent sequence boundary. Resetting only one leaves hidden State or eligibility traces carrying over from the previous trajectory.

Yes, but use exactly one mapping owner: either create one brainstate.nn.Map and construct the algorithm around it, or pass an unmapped model to braintrace.compile(..., vmap=True). Never pass an already mapped model to compile with vmap=True, which would map it twice.