OUR APPROACH

Memory Poisoning Detection via Graph Topology

DEPLOYMENT STATUS:RESEARCH · PLANNED FEATURE
SECTION 1 — OVERVIEW

Overview & Core Concept

Our planned Vector SIEM feature detects MINJA-style memory poisoning through graph topology analysis. It tracks embedding clusters over time, detects anomalous cluster formation, and re-scores memory at retrieval. The approach is informed by published work in structural anomaly detection for agent memory, including MemAudit and MemSentry.

SECTION 2 — METHODOLOGY

Implementation Architecture

  • 01Graph topology analysis of vector store: Evaluates edge densities and shortest-path neighborhood metrics across vector records rather than relying solely on naive cosine distance.
  • 02Embedding cluster tracking over time: Maintains sliding topological windows to detect gradual, slow-poisoning drifts across multi-turn agent interactions.
  • 03Anomalous cluster formation detection: Flags suspicious dense island subgraphs that form in uncharacteristic regions of the semantic space.
  • 04Retrieval-time re-scoring: Dynamically applies topological penalty multipliers to candidate memories before injecting them into model context.
  • 05Provenance tracking (who inserted what, when): Immutable audit trail binding every memory write to authenticated session ID, tool call trace, and timestamp.
SECTION 3 — LIMITATIONS

Honest Boundary Disclosures

Memory poisoning is an active research area. We implement the approach and measure our results under a documented, reproducible protocol. We do not claim to have solved the problem.

// Informed by MemAudit (arXiv:2605.23723), MemSentry (arXiv:2609.08747), and OWASP ASI06.