| _version_ | 1866901537013366784 |
|---|---|
| author | Singh, Shivam |
| author_facet | Singh, Shivam |
| contents | <p>This paper introduces Instance-Based Memory Augmentation (IBMA), a memory framework designed to solve the "Contextual Amnesia" and memory bloat issues in Large Language Models (LLMs). Traditional RAG systems suffer from linear memory growth, eventually leading to hardware failures (OOM) in resource-constrained environments.</p> <p> </p> <ul> <li> <p>Metabolic Lifecycle: Introduces a dynamic graph architecture where nodes possess a decaying Vitality Score (V) and a reinforcement-based Lignification Weight (W)</p> </li> <li> <p>Memory Homeostasis: Implements a periodic "Slashing" cycle that prunes low-utility nodes, allowing the system to maintain a stable memory footprint indefinitely.</p> </li> <li> <p>Contextual Frames: Utilizes metadata tagging on relational edges to allow LLMs to switch between different worldviews or chronological states without catastrophic forgetting.</p> </li> <li> <p>Our empirical results, conducted on an HP ProLiant enterprise server, demonstrate that IBMA successfully eliminates context amnesia while maintaining a constant asymptotic ceiling on memory consumption. </p> </li> <li> <p>Keywords: LLM Memory, GraphRAG, Contextual Amnesia, Bio-mimetic AI, Resource-efficient Machine Learning.</p> </li> </ul> <div> </div> <p> </p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_18775789 |
| institution | Zenodo |
| language | |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | Lignified Memory Graphs: Instance-Based Architecture for Mitigating Contextual Amnesia in Large Language Models Singh, Shivam <p>This paper introduces Instance-Based Memory Augmentation (IBMA), a memory framework designed to solve the "Contextual Amnesia" and memory bloat issues in Large Language Models (LLMs). Traditional RAG systems suffer from linear memory growth, eventually leading to hardware failures (OOM) in resource-constrained environments.</p> <p> </p> <ul> <li> <p>Metabolic Lifecycle: Introduces a dynamic graph architecture where nodes possess a decaying Vitality Score (V) and a reinforcement-based Lignification Weight (W)</p> </li> <li> <p>Memory Homeostasis: Implements a periodic "Slashing" cycle that prunes low-utility nodes, allowing the system to maintain a stable memory footprint indefinitely.</p> </li> <li> <p>Contextual Frames: Utilizes metadata tagging on relational edges to allow LLMs to switch between different worldviews or chronological states without catastrophic forgetting.</p> </li> <li> <p>Our empirical results, conducted on an HP ProLiant enterprise server, demonstrate that IBMA successfully eliminates context amnesia while maintaining a constant asymptotic ceiling on memory consumption. </p> </li> <li> <p>Keywords: LLM Memory, GraphRAG, Contextual Amnesia, Bio-mimetic AI, Resource-efficient Machine Learning.</p> </li> </ul> <div> </div> <p> </p> |
| title | Lignified Memory Graphs: Instance-Based Architecture for Mitigating Contextual Amnesia in Large Language Models |
| url | https://doi.org/10.5281/zenodo.18775789 |