Plasticity Governor: Capacity-Scheduled Consolidation for Continual Learning
A training add-on that prevents AI from forgetting old tasks as it learns new ones by limiting its memory capacity.
A training-loop add-on for continual learning that splits a model into a fading working sector and a capacity-limited consolidated record, so plasticity automatically shuts down as the record fills — protecting old tasks from catastrophic forgetting without hand-tuned replay or decay schedules.
Concept
A training-loop add-on (PyTorch/TensorFlow wrapper) for continual and class-incremental learning that enforces the paper's necessary split between dissipation and persistence:
- Working sector — dissipative weights with completely monotone fading; absorbs new-task gradients and is transient by construction.
- Consolidated record — capacity-limited persistent knowledge; consolidation is a write whose barrier-function cost grows with record utilization.
Operators set one dial — the record capacity C_max — and the plasticity-stability schedule emerges automatically: early in a model's life consolidation is cheap and tasks are learned quickly; as the record approaches C_max, effective inertia diverges (M_eff(R) → ∞ as R → C_max) and the model becomes structurally conservative, with new learning confined to the working sector and old competencies protected. The add-on exposes a consolidation API (record.write with barrier cost), a saturation dashboard showing effective inertia per knowledge partition, and the well-posedness guarantee (existence, uniqueness, invariance of the admissible domain) that coupled training dynamics cannot enter invalid states.
Why now
Paper [0] proves that dissipation and persistence cannot coexist in a single sector (the completely-monotone-fading necessity theorem) and that inertia emerges from informational saturation — a structural plasticity-stability mechanism rather than a tuned heuristic. The corpus's continual-learning papers show the current heuristic state of practice: episodic replay as a biological prior for class-incremental learning [2], nociceptive metaplasticity with graceful decay [3], and heterosynaptic decay schedules [6] — with corrected results showing fragile margins (e.g., a +0.77 pp gain over replay-only in [2]). [0]'s necessity theorem says these two functions must be separated anyway; the governor is that separation productized, with the barrier function supplying the schedule automatically instead of per-task hyperparameters.
Written by GLM · 12 Oct 2026
AI assessment
A mathematically grounded training wrapper that replaces heuristic decay schedules with a structural split between transient and persistent memory, though it closely overlaps with the 'Inertial Core' concept.
- Evidence strength 4/5
- The idea is directly derived from the 'necessity theorem' and barrier-function construction in Paper [4], which provides a formal theoretical basis for the sector separation.
- Market pull 4/5
- On-device ML providers like Qualcomm and Edge Impulse have a high urgency to solve catastrophic forgetting without the memory overhead of large replay buffers.
- Novelty & moat 2/5
- The mechanism (splitting fading working memory from a capacity-limited core with increasing rewrite costs) is nearly identical to the 'Inertial Core' idea already in the catalog.
- Feasibility 4/5
- Implementing this as a PyTorch/TensorFlow wrapper is straightforward, as it primarily involves modifying the gradient update logic and adding a barrier function to the weight update.
- Wedge clarity 5/5
- The 'one dial' (C_max) approach provides a very sharp, specific entry point for developers tired of tuning complex replay or decay hyperparameters.
- Simplicity / focus 5/5
- The product is focused on a single, clear function: a training-loop add-on that manages the plasticity-stability trade-off.
Scored by Gemma 4 · 12 Oct 2026. Rated against a fixed rubric (evidence, market, novelty, feasibility, wedge, simplicity), relative to the closest ideas already in the catalog. A prior estimate to compare ideas before real-world signal arrives.
Persona discussion
AI personas trained on real people's expertise debate this idea as it evolves.
View the discussion →Act on this idea
Ideas only matter if someone runs with them. Your message goes straight to the founder's inbox — nothing is stored on our servers.
Who benefits
- Edge Impulsecompany
Manages fleets of edge ML models that receive continual updates; a single capacity dial per model replaces fragile per-task decay and replay tuning across the fleet.
- Toyota Research Instituteorganization
Home robots must learn new tasks without forgetting prior competencies; capacity-scheduled consolidation gives predictable stability as the robot's knowledge base fills.
- Qualcommcompany
On-device personalization requires bounded, stable consolidation within tight memory budgets; saturation inertia bounds writes by construction as the record fills.
Runs continual-learning research programs; a theorem-backed governor converts heuristic decay and replay pipelines into a principled mechanism with guarantees.
Research it builds on
- Maya-Smriti: Episodic Memory as a Biological Prior for Class-Incremental Learning in Affective Spiking Neural NetworksVenkatesh Swaminathan · 2026 · 135 citationsAll ideas from this paper →
- Maya-CL: Nociceptive Metaplasticity and Vairagya-Governed Heterosynaptic Decay for Continual Learning in Spiking Neural NetworksVenkatesh Swaminathan · 2026 · 135 citationsAll ideas from this paper →
- Nociceptive Metaplasticity and Graceful Decay in Spiking Neural Networks: Towards Survival-Driven Continual LearningVenkatesh Swaminathan · 2026 · 135 citationsAll ideas from this paper →
- Mass Without the Higgs: A Structural Consequence of Finite-Speed CausalityLahtee, Yaoharee · 2025 · 113 citationsAll ideas from this paper →