[CVPR 2026] GAM-RAG: From Static Search to Evolutionary Retrieval via Gain-Adaptive Memory

Executive Summary

TL;DR: GAM-RAG is a training-free framework that transforms RAG from a stateless search process into a dynamic, "plastic" memory system. Inspired by cognitive neuroscience, it learns from every retrieval episode to build "logical shortcuts," eventually improving accuracy by 8.19% while slashing inference costs by over 60%.

Background Positioning: While most RAG research focused on better indexing (GraphRAG) or better retrieval (Hybrid RAG), GAM-RAG tackles the statelessness of the retrieval process itself. It sits at the intersection of Knowledge Graphs and Online Continual Learning, moving the field toward "Long-term Memory" for LLMs.

Problem & Motivation: The "Stateless" Retrieval Trap

Current RAG implementations—even advanced graph-based ones—treat every query as a "first-time" event. If you ask a multi-hop question today and a similar one tomorrow, the system repeats the same expensive graph traversal from scratch.

The authors identified two major pain points:

1. Redundant Reasoning: Static indices cannot capture the "provenance" of successful past reasoning paths.
2. Feedback Noise: Simply updating memory after every query is dangerous; noisy feedback from a misaligned LLM judge can "poison" the global index.

Methodology: The Core of GAM-RAG

GAM-RAG replaces the rigid, relation-heavy KG with a lightweight hierarchical graph (Entities $\to$ Sentences $\to$ Passages). The breakthrough lies in how this graph evolves.

1. The Multi-Channel Memory

Every sentence node maintains two memory vectors:

• ${\mathbf{m}}_{task}$: Captures semantic relevance to historical queries.
• ${\mathbf{m}}_{time}$: Specifically manages temporal constraints—a classic weakness of semantic search.

2. Kalman-Inspired Gain Updates

This is the "secret sauce." Instead of a fixed learning rate, GAM-RAG treats memory as a dynamic state with Uncertainty (Perplexity $\pi$).

When the LLM judge provides feedback ($y_i$), the system calculates a Kalman Gain ($K_i$): $K_t=\frac{{\pi }_t}{{\pi }_t+R}$

• If Memory is Uncertain ($\pi$ is high): $K\approx 1$, leading to a Fast Warm-up (rapidly absorbing new knowledge).
• If Memory is Stable ($\pi$ is low): $K$ decreases, leading to Damped Refinement (protecting the memory from recent noise).

3. Memory-Guided Propagation

During inference, these memories don't just sit there; they act as priors. They re-weight the graph propagation, effectively "pulling" supportive evidence closer in the representation space and allowing the system to skip irrelevant branches.

Experiments & Results: Efficiency Meets Accuracy

The authors tested GAM-RAG across diverse benchmarks, including 2WikiMultiHop, HotpotQA, and TimeQA.

Performance Gains

The results show a clear "warm-up" effect. As the system "experiences" more queries (turns), the accuracy climbs consistently.

Efficiency Breakthrough

Most impressively, the memory mechanism allows the system to find the right answer faster. By turn 5, the online latency dropped by 61% because the "memory-aligned priors" guided the search directly to the relevant evidence, avoiding exhaustive graph traversal.

Figure: Ablation under long-term memorization shows the stability of the full model (Full) versus versions without task/time memory.

Critical Analysis & Conclusion

Takeaway

GAM-RAG proves that RAG does not need to be static. By treats retrieval residues as valuable training signals, we can create a system that gets smarter and cheaper the more it is used. The use of Kalman filtering to balance "stability vs. plasticity" is a masterclass in applying classical control theory to modern AI.

Limitations

• Feedback Dependency: The system relies on an "LLM-as-a-judge" for the signal $y_i$. If the judge is consistently biased, the memory will stabilize on wrong information.
• Memory Scaling: While the index is lightweight, maintaining two vectors and perplexity for every sentence in a massive corpus may eventualy pose storage challenges.

Future Outlook

This work opens the door for Cross-User Experience Sharing where a RAG system deployed in a corporate environment could learn from one department's queries to benefit another, creating a truly evolving "Corporate Brain."