SALM: Adaptive AI Long-Term Memory Learns What to Remember, Retrieve, and Forget

Paper Overview

The survey Human-Inspired Perspectives: A Survey on AI Long-Term Memory is published in Proceedings of the IEEE . The collaboration involves Memories.ai, Hong Kong University of Science and Technology (Guangzhou), University of Cambridge, University of Bristol, Columbia University, and Meta researchers. First author Zihong He is a PhD student at HKUST Guangzhou; corresponding author Weizhe Lin (Cambridge PhD) works on multimodal retrieval-augmented QA; co-corresponding author Junxiao Shen (Cambridge PhD, former Meta) co-founded Memories.ai.

Memory Types and Storage Forms

The paper maps human long-term memory categories to AI systems:

Episodic memory : interaction records and historical events.

Semantic memory : knowledge bases and learned concept relationships.

Procedural memory : policies from reinforcement learning, adjustable rules, and code.

Storage is classified by location:

Parametric memory : information encoded in model parameters.

Non-parametric memory : information kept in external stores (databases, file systems). RAG belongs here.

These two dimensions are orthogonal; e.g., semantic knowledge can reside in parameters or an external knowledge base.

SALM Cognitive Architecture

SALM (Self-Adaptive Long-term Memory) extends classic cognitive architectures (ACT-R, Soar, Sigma, CoALA) with four modules:

Perception register

Working memory

Long-term memory (both parametric and non-parametric)

Action execution

Key innovation: learnable adapters make storage, retrieval, and forgetting decisions based on task feedback. The adapters choose trigger timing, memory type, and specific content. Adapter optimization itself accumulates procedural memory.

Adapter Mechanisms

Implementation Backbone

Meta-Llama-3-8B-Instruct serves as the base model. Parametric memory uses LoRA adapters; non-parametric memory uses external context.

Storage Adapter — What to Remember

Triggered each session. A discriminator network first decides whether to save, then selects memory type. For parametric storage, the system chooses between reusing an existing LoRA adapter or creating a new one, guided by quality scores, remaining capacity, and semantic branching. Design goals include filtering low-quality data, poisoning inputs, and privacy-sensitive information.

Retrieval Adapter — How to Find

Given current context, the adapter picks memory sources. For non-parametric memory, it scores joint features of memory entries and query, selects top-K candidates, then ranks by relevance and memory quality. For parametric memory, a routing network weighs LoRA storage quality, historical call frequency, and relevance to the query to choose which adapter to invoke.

Forgetting Adapter — When to Forget

Decides which external entries or LoRA adapters to remove. Active forgetting frees capacity and reduces retrieval interference. Implementation: first judge whether to trigger forgetting, then pick memory type, finally score candidates for removal. Both memory types use learnable networks for selection.

Training Memory Strategies with Task Feedback

A multi-branch decision tree explores combinations of storage, retrieval, and forgetting actions. Each dialogue history maintains up to 20 active paths. Each path generates an answer; reward combines answer quality, improvement over a no-memory baseline, and the effectiveness of storage, forgetting, and parametric-memory training. Policy gradient updates the memory-handling strategies.

Long-Dialogue Experiments

Dataset : LongMemEval_S — 500 long dialogue histories, each ~115k tokens, with evaluation questions requiring cross-session information (temporal reasoning, knowledge update).

Protocol : Each round randomly shuffles data, samples 128 histories, runs 4 independent trials, tracks cumulative average reward.

Comparisons :

Dual memory (parametric + non-parametric) vs. non-parametric only vs. parametric only.

Full SALM (with adapters) vs. no-adapter baseline (save all sessions, load all non-parametric memory into context, call largest LoRA adapter).

Results : SALM with adapters shows a more stable upward trend in cumulative average reward. Dual-memory configuration gradually stabilizes; non-parametric-only rises; parametric-only shows only slight early improvement. Adapters exhibit a more pronounced learning effect on non-parametric memory handling.

Code is publicly available.