RAMPART: Registry-based Agentic Memory with Priority-Aware Runtime Transformation
RAMPART optimizes LLM agent memory with priority-aware runtime transformation, enhancing task success.
Key Findings
Methodology
RAMPART is a registry-based memory model utilizing five composable primitives (promote, gate, write, evict, rollback) to manipulate named addressable blocks before compilation. It employs priority-aware context compilation to control the absolute position of blocks and their structural relationship with task queries.
Key Results
- Experiments on Qwen3-8B Q4 show task success drops sharply after the seventh block position, with block grouping enhancing success rates.
- Cross-model replication shows content-priming effects across models, with block grouping increasing Mistral's success rate fivefold.
- Relevance gating reduces prompt cost by 67.8% while recovering 83% of the success rate.
Significance
RAMPART significantly improves task success for LLM agents by controlling memory block positions and order at compile-time. This approach addresses traditional memory model issues in context assembly, reducing prompt costs and demonstrating consistent performance improvements across multiple models.
Technical Contribution
RAMPART introduces a novel memory model allowing fine-grained control over memory blocks at compile-time. Its priority-aware context compilation and block grouping strategies have been validated across multiple models, offering new engineering possibilities.
Novelty
RAMPART is the first to treat memory block positions and task query relationships as programmable variables, significantly enhancing task success compared to existing methods.
Limitations
- RAMPART may face performance bottlenecks when handling very large registries.
- Further validation is needed across different models and tasks.
Future Work
Future research could explore RAMPART's application in larger-scale models and multi-task environments, as well as optimize its performance on different hardware.
AI Executive Summary
RAMPART introduces a novel registry-based memory model that optimizes LLM agent task success through priority-aware runtime transformation. Traditional memory models face issues with fixed order and content in context assembly, leading to decreased task success. RAMPART addresses these issues by using five composable primitives (promote, gate, write, evict, rollback) to manipulate named addressable blocks before compilation, achieving priority-aware context compilation.
Experimental results show that RAMPART significantly improves task success on Qwen3-8B Q4, especially under block grouping strategies, with success rates increasing by tens of percentage points. Cross-model replication confirms its consistency across different models, with block grouping increasing Mistral's success rate fivefold. Relevance gating reduces prompt cost by 67.8% while recovering 83% of the success rate.
RAMPART's innovation lies in treating memory block positions and task query relationships as programmable variables for the first time, offering new engineering possibilities. However, it may face performance bottlenecks when handling very large registries. Future research could explore its application in larger-scale models and multi-task environments.
Deep Analysis
Background
As LLMs are applied to various tasks, optimizing memory models becomes crucial. Traditional memory models face issues with fixed order and content in context assembly, leading to decreased task success. RAMPART offers a new approach to memory management through priority-aware runtime transformation.
Core Problem
Traditional memory models face issues with fixed order and content in context assembly, leading to decreased task success. Optimizing memory block positions and order at compile-time is a significant research challenge.
Innovation
RAMPART uses five composable primitives (promote, gate, write, evict, rollback) to manipulate named addressable blocks before compilation, achieving priority-aware context compilation. Its innovation lies in treating memory block positions and task query relationships as programmable variables.
Methodology
- �� Use five composable primitives to manipulate memory blocks.
- �� Priority-aware context compilation controls memory block positions and order.
- �� Block grouping strategy enhances task success rates.
Experiments
Experiments were conducted on Qwen3-8B Q4, testing the impact of different memory block positions and orders on task success. Cross-model replication confirmed its consistency across different models.
Results
RAMPART significantly improves task success on Qwen3-8B Q4, especially under block grouping strategies, with success rates increasing by tens of percentage points. Cross-model replication confirms its consistency across different models.
Applications
RAMPART can be used to optimize memory management for LLM agents, enhancing task success rates and reducing prompt costs, applicable to various tasks requiring efficient memory management.
Limitations & Outlook
RAMPART may face performance bottlenecks when handling very large registries. Future research could explore its application in larger-scale models and multi-task environments.
Plain Language Accessible to non-experts
Imagine a kitchen where RAMPART acts as a smart assistant, helping the chef prepare all necessary ingredients and tools before cooking. Traditional assistants might place everything on the table at once, but RAMPART prioritizes the most important items, placing them within easy reach. This way, the chef can complete each dish faster and better. RAMPART optimizes the entire cooking process, reducing unnecessary waste and time.
ELI14 Explained like you're 14
Hey there! Imagine you're playing a fast-paced game. RAMPART is like your super helper, getting all the tools you need ready before the game starts. Traditional helpers might throw everything at you at once, but RAMPART prioritizes the most important tools, putting them where you can easily grab them. This way, you can win the game faster and better! Isn't that cool?
Glossary
RAMPART
A registry-based memory model optimizing LLM agent task success.
Used to control memory block positions and order.
Priority-aware
Adjusting memory block positions based on task importance.
Used in RAMPART to optimize context compilation.
Block grouping
Combining related memory blocks to enhance task success.
Validated in experiments for its effectiveness.
Relevance gating
Filtering irrelevant content to reduce prompt cost.
Used in RAMPART to optimize memory management.
Compile-time
The moment before task execution when memory blocks are optimized.
RAMPART optimizes memory blocks at compile-time.
Open Questions Unanswered questions from this research
- 1 How can RAMPART's performance be optimized in larger-scale models?
- 2 What is RAMPART's performance in multi-task environments?
Applications
Immediate Applications
LLM Agent Optimization
RAMPART can optimize memory management for LLM agents, enhancing task success rates.
Long-term Vision
Intelligent Systems
RAMPART's technology could be applied to broader intelligent systems, improving their efficiency and performance.
Abstract
RAMPART is a compile-time memory model and pure in-RAM block registry for LLM-based agents. Context assembly is a programmable runtime operation where content is compiled from a structured registry under explicit policy for ordering, inclusion, and eviction. Five composable primitives (promote, gate, write, evict, rollback) act on named addressable blocks before compilation at zero prompt-token cost. Provenance tags and non-evictable authorship flags implement a permissioned memory model with block-level ownership. Controlled probes with Qwen3-8B Q4 show that compile-time placement and the structural relationship between blocks and the task query affect task success, with the cliff falling at roughly the seventh block position when the task follows the registry and the twelfth when it precedes. Grouping the critical block with content-adjacent neighbours and promoting the group as a unit lifts task success by tens of percentage points at positions where single-block placement fails. Cross-model replication on Qwen2.5-7B, Llama-3.1-8B, Mistral-7B-v0.3, and Qwen3-14B shows the content-priming effect appears at the same absolute positions across families, with magnitude varying with model strength. Block grouping raises Mistral's mean pass rate roughly fivefold at the hardest registry size, and a smaller model with the intervention can outperform a larger model without it in the mid-registry zone. Relevance gating reduces prompt cost by 67.8\% while recovering 83% of the promoted-condition success rate. Schema eviction produces 0% invocations against 100% with the schema present, a property policy-based approaches cannot guarantee by construction. Shared-registry coordination reduces inter-agent communication to a method call at zero coordination token cost.