Grafy News
Try Grafy free

New Research Proposes Parametric Memory That Can Forget

Researchers have introduced a system that manages the lifecycle of information stored in model weights, allowing transient facts to decay naturally.

Illustration for the news story: New Research Proposes Parametric Memory That Can ForgetSource: Grafy Research Lab

Large language models often struggle to balance the need to remember permanent facts with the necessity of forgetting temporary information. New research from Samsung Research America suggests that instead of trying to retain all knowledge indefinitely, models should use a managed parametric memory system that allows certain information to decay over time.

Why do current AI memory systems struggle?

Most current approaches to long-term AI memory fall into two categories: Retrieval-Augmented Generation (RAG) and Continual Learning (CL). According to the paper "Some Memories Must Fade: Persistence and Decay of User Memories in Model Parameters," RAG stores information externally and injects it into the prompt, while CL attempts to bake knowledge directly into the model's weights.

However, the authors note that CL is fundamentally misaligned with how human memory works. CL typically focuses on the objective of retaining information indefinitely to prevent "catastrophic forgetting." This is insufficient for a personalized assistant. For example, a model should remember that a user's profession is a teacher for years, but it should not remember that the user ate a bagel for breakfast several months later. Existing CL methods often fail to represent outdated information because they are designed to store values well rather than manage their expiration.

How does the proposed memory substrate work?

To solve this, the researchers proposed a system that separates storage dynamics from the decision-making policy. The first component is a "memory substrate," which serves as the physical storage mechanism. This substrate uses a fixed-capacity LoRA (Low-Rank Adaptation) adapter.

To prevent different pieces of information from interfering with one another, the system routes each fact to a sparse kk-hot combination of adapter ranks. This isolation is crucial for stability. Furthermore, the researchers implemented an independent temporal decay rate for each rank. This mechanism allows the model to undergo a process of natural forgetting—where outdated information fades away—without requiring explicit, computationally expensive unlearning processes. This approach enables the model to maintain a balance between long-term stable facts and transient, short-lived observations.

What role does the learned policy play?

While the substrate handles the "how" of storage, a learned policy handles the "what" and "how long." The researchers trained a lightweight controller with 2 million parameters to act as a manager for incoming user events. This controller performs three specific actions for every new piece of information: it decides whether to write the data (a binary gate), how deeply to write it into the model's parameters (depth), and how long the information should remain valid (lifespan).

Stable, recurring information is assigned deep, durable writes. In contrast, transient events are assigned short or no lifespans. By using this controller, the system can react to changing user beliefs. If a user provides contradictory information, the policy can decide to overwrite or weaken the previous belief, ensuring the model's internal knowledge reflects the user's current reality.

How effective is this memory management?

To test this system, the authors created a dataset consisting of 909,909 controlled, long-horizon user trajectories. These trajectories included various scenarios such as belief revision, time-based expiry, and reinforcement of routines. The researchers evaluated the system based on its ability to recall valid information and its ability to "abstain" (refrain from answering) once information had expired.

The results showed that the learned management system achieved 56% class-balanced accuracy. This outperformed the strongest CL baseline, which reached 42%, and a compute-matched retrieval-based memory manager, which reached 50%. The researchers noted that their system recovered between 86% and 92% of the performance of an "oracle" (a perfect system with total knowledge of when information expires), demonstrating that effective lifecycle decisions can indeed be inferred from online observations.

How can researchers apply these concepts?

For researchers working on the frontiers of agentic workflows and long-term personalization, this work highlights the importance of moving beyond simple retrieval. Understanding how to manage the lifecycle of data—knowing when to reinforce a fact and when to let it decay—is essential for creating assistants that feel truly intuitive and context-aware.

While current commercial models primarily rely on context windows and RAG, the shift toward managed parametric memory could eventually lead to models that require much smaller prompts because their internal weights are more accurately tuned to the user's evolving state.

If you are currently exploring complex AI research papers or technical documentation regarding model architectures and memory management, you can use Reader to perform deep research over PDFs and academic papers, allowing you to query specific technical details and mathematical formulations within the text.

Frequently asked questions

Does the proposed system allow models to forget information?

Yes, the system is designed to allow transient information to fade. By using a memory substrate with independent temporal decay rates on each rank, the model can let outdated or temporary facts decay naturally over time without the need for manual unlearning or explicit deletion commands.

How does the controller decide what to remember?

The 2-million-parameter controller makes decisions based on incoming user events. It determines three things: whether to perform a write operation, the depth of the gradient write to ensure durability, and the specific lifespan assigned to that piece of information based on its perceived stability or transience.

Why is continual learning considered insufficient for this task?

Continual learning typically focuses on preventing the loss of old knowledge when learning new things, effectively aiming to retain all information indefinitely. The researchers argue this is insufficient because a functional memory system must also be able to expire outdated information and update beliefs when they are contradicted.

Source: Sanidhya Vijayvargiya, Rahul Lokesh, "Some Memories Must Fade: Persistence and Decay of User Memories in Model Parameters"


Source: Grafy Research Lab — Some Memories Must Fade: Persistence and Decay of User Memories in Model Parameters