Decoupling Stability and Plasticity via Dual Memory Architectures

749 concepts490 papers4 proposals

When a neural network learns a new task or adapts to a novel environment, it faces a fundamental dilemma: stability versus plasticity. Plasticity demands that the network flexibly alter its weights to absorb new information. Stability requires that the network preserve the precise weight configurations responsible for historical knowledge.

In standard deep learning architectures, these two forces are in direct conflict. When gradients from a new data distribution aggressively update a single set of shared weights, they inevitably overwrite previously learned features—a phenomenon known as catastrophic forgetting.

To resolve this tension, a powerful paradigm has emerged: decoupling the learning process into two distinct, interacting tracks. Heavily inspired by biological complementary learning systems—where the brain utilizes the hippocampus for rapid, short-term acquisition and the neocortex for slow, structured consolidation—this approach splits the responsibility of learning. By isolating immediate updates from the permanent knowledge base, models can aggressively adapt without destroying their past.

Dual-Track Parameter Architectures

One way to implement this decoupling is directly within the network's parameter space, maintaining separate sets of weights optimized for different temporal scales.

Instead of forcing a single model to find an impossible compromise in the loss landscape, researchers have proposed maintaining parallel models that track updates at different speeds. For instance, Arani et al. maintain two auxiliary models derived from the primary network using exponential moving averages at different decay rates. A fast-decaying model captures recent updates to maintain high plasticity, while a slow-decaying model preserves long-term structural knowledge to ensure stability. Similarly, in meta-learning contexts, Caccia et al. utilize a dual-weight architecture that maintains fast weights for immediate task adaptation alongside slow weights that accumulate long-term knowledge across tasks.

                  ┌────────────────────────┐
                  │   Incoming New Data    │
                  └───────────┬────────────┘
                              │
              ┌───────────────┴───────────────┐
              ▼                               ▼
   ┌────────────────────┐          ┌────────────────────┐
   │   Plastic Track    │          │   Stable Track     │
   │ (Fast Weights/EMA) │          │ (Slow Weights/EMA) │
   └──────────┬─────────┘          └──────────┬─────────┘
              │                               │
              └───────────────┬───────────────┘
                              ▼
                  ┌────────────────────────┐
                  │ Consolidated Knowledge │
                  └────────────────────────┘

Alternatively, the network can be structurally split into static and plastic branches. Zhang et al. implement a dual-stream backbone that combines a fixed, pre-trained feature extractor with a fully learnable branch, blending their outputs via bilinear pooling to preserve robust base representations while adapting to new distributions. Taking this isolation a step further, Lin et al. optimize two entirely separate parameter sets for each task—one focused purely on stability and another on plasticity—before merging their specialized weights into a single network. This dual-track approach ensures that the model can freely explore new regions of the loss landscape without immediately corrupting historical decision boundaries.

Dual-Tier Memory Systems for Agents and Sequences

The philosophy of decoupling stability and plasticity extends beyond weight optimization into context management and data storage. When processing long video sequences or acting as autonomous agents in changing environments, models face strict context windows and memory constraints. Forcing a single memory buffer to hold everything leads to information clutter and the rapid eviction of critical historical data.

To manage this, modern architectures split data storage into a short-term episodic buffer and a long-term semantic archive. In long video understanding, Song et al. store recent, dense visual tokens in a first-in-first-out short-term buffer while archiving summarized, sparse tokens in long-term storage. This allows the model to maintain immediate local context alongside a broad global understanding of extended sequences. For multimodal tracking, de Melo stores recent multimodal embeddings in a short-term buffer while maintaining long-term knowledge via a pool of learnable parameters, and Feng utilizes a dual-memory framework to balance rapid learning from recent frames with a persistent understanding of a target's identity.

For personalized and reasoning agents, this manifests as a split between raw experiences and abstract profiles:

  • Episodic/Short-Term Layer: Captures raw, time-stamped interaction sequences and immediate situational awareness.
  • Semantic/Long-Term Layer: Distills these raw events into stable, high-level world models, user preferences, and behavioral traits.

As demonstrated by Anokhin et al., integrating structured semantic triplets with raw episodic observations allows agents to maintain a high-level world map while retaining fine-grained historical context. Similarly, Zhang et al. use this dual-layer structure to evolve user personas at test time, ensuring that noisy or anomalous short-term interactions do not immediately distort the agent's core understanding of the user.

The Consolidation Pipeline

Decoupling the acquisition of knowledge from its retention is only half the solution; the system must eventually bridge the gap between the two tracks. This is achieved through a structured consolidation pipeline, where short-term, plastic updates are safely distilled into the stable, long-term core.

Liang et al. show that a two-stage learning framework—which completely separates the acquisition of new knowledge from the preservation of existing knowledge—avoids the inherent competition present in single-objective optimization. In reinforcement learning, Zhou et al. separate rapid online interaction from a stable offline distillation phase. This allows an agent to optimize aggressively for new tasks online without the immediate constraint of preserving old skills, which are consolidated later from stored experiences.

┌────────────────────────┐      Distillation      ┌────────────────────────┐
│  Active Plastic Head   │───────────────────────>│  Frozen Historical Core│
│ (Aggressive Alignment) │                        │  (Knowledge Retention) │
└────────────────────────┘                        └────────────────────────┘

This consolidation often relies on structural re-parameterization or multi-head classification to prevent interference. Nie et al. decouple parameters into interacting short-term and long-term modules with different optimization goals, ensuring high representational capacity for new tasks while maintaining a stable, class-balanced historical representation. By routing new data through plastic heads or temporary network extensions—as explored by Gao et al. and Silver et al.—the system borrows features for immediate learning while maintaining the absolute integrity of its permanent memory. Ultimately, by separating the how to learn from the how to remember, dual-memory architectures provide a robust blueprint for truly continuous learning.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_complementary_learning_systems_eb9feed4,
  title = {complementary learning systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/complementary-learning-systems/rdr_eb9feed4fc0e437ea7b50b1e90347490},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_complementary_learning_systems_eb9feed4_v1,
  title = {complementary learning systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/complementary-learning-systems/rdr_eb9feed4fc0e437ea7b50b1e90347490/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake