What question did this study set out to answer?

This research aims to address structural failures in stateless RAG systems through a novel memory architecture.

April 28, 2026Open Access

ContextForge: Architecture-First Persistent Agentic Memory via Model Context Protocol — Multi-Trigger OR-Gate Security, OR-Set CRDT Sync, Tri-Core Failover, Differential Context Injection, Recency-Weighted Retrieval, and a Memory Quality Benchmark

Key Points

This research aims to address structural failures in stateless RAG systems through a novel memory architecture.
Developed a five-pillar agentic memory architecture with a 22-tool Model Context Protocol (MCP).
Conducted benchmark tests (n = 300) to evaluate memory integrity, security, and retrieval methods.
Implemented two operating modes in the ReviewerGuard v3 OR-Gate for detailed experiments.
Achieved an Attack Block Rate of 55% with a benign false positive rate of 1%, an improvement from 25%.
Memory Integrity Score increased from 0.229 to 0.600 via recency-weighted retrieval techniques.
Composite Safety Index was reported at 79.7%, indicating overall enhanced system resilience.

Abstract

ContextForge v2. 3 — Nexus MCP Server: Supplementary Materials ContextForge is a five-pillar agentic memory architecture that closes three structural failure modes of stateless RAG systems: zero adversarial block rate, high failover latency, and unbounded context injection. It exposes all capabilities as a 22-tool Model Context Protocol (MCP) server — delivering persistent, adversarially-hardened memory across Claude Desktop, Cursor, VS Code, and Windsurf with zero client-side changes. This record contains the supplementary data, benchmark results, figures, and source code for the paper "ContextForge: Architecture-First Persistent Agentic Memory via Model Context Protocol — Multi-Trigger OR-Gate Security, OR-Set CRDT Sync, Tri-Core Failover, Differential Context Injection, Recency-Weighted Retrieval, and a Memory Quality Benchmark". Key Results Security / Injection Defense (Suite 14, n = 300) Attack Block Rate (ABR): 55% at benign FPR = 1% (vs 25% FPR in v1 PAPER mode, −24 pp) Edge-case FPR: 16% (vs ~97% in v1 PAPER mode, −81 pp) External validation macro-F1: 75. 5% (deepset/prompt-injections dataset) Composite Safety Index Φ = 79. 7% Memory Integrity (Suite 15 v2, n = 160, 6-system comparison) Memory Integrity Score (MIS): 0. 8012 — mean (Recall@3, UpdateAccuracy, DeleteAccuracy, PoisonResistance) Update accuracy: 0. 229 → 0. 600 (+37. 1 pp) via recency-weighted BM25 (score = BM25 × exp (−λ·age), λ = 0. 0001 s⁻¹) Operating Modes The ReviewerGuard v3 OR-Gate ships two modes selectable via CFMODE: paper (default) — word-level entropy H* = 3. 5, exact paper reproduction experiment — char-level entropy H* = 4. 8, soft Pass 2, auto-perplexity gate (P* = 231. 8), reducing edge-case FPR from ~97% to 16% Archive Contents benchmarkᵣesults. zip — All benchmark JSON outputs (Suites 7–12, 14, 15; security comparison tables) figures. zip — Publication figures (PNG) + Python generator scripts for full reproducibility sourcecode. zip — Core Python modules, LaTeX paper source, and bibliography Reproducibility pip install -r requirements. txt # Full v5. 0 benchmark suite (375 tests) python -X utf8 benchmark/testᵥ5/runₐll. py # Suite 14 — FPR fix evaluation (300 samples × 5 baselines) python -X utf8 benchmark/suites/suite₁4fprfixₑval. py # Regenerate all paper figures python research/figures/genₐll. py License: MITRepository: https: //github. com/parnish007/contextforge

Read Full Paperexternally

KI fragen

Bookmark

View Full Paper