Memory Graph connects every attack, model, policy rule and calibration run, so each scan learns from the last.
What Memory Graph does
Most security scans end when the report is delivered. The next scan starts from zero, and nobody can say whether an attack that worked last month is a variant of one that worked today. Memory Graph is the memory of AIZA-HexTyx. It stores what every scan learned and connects it, so each scan is sharper than the last.
Scan findings→Memory Graph nodes and edges→Clusters and transfers→Aegis Runtime Shield calibration→AXIOM explains
Problems it solves
Problem
How Memory Graph addresses it
Every scan starts from zero and nothing carries over
Each attack attempt, model, policy rule and calibration run is written as a node and linked, so intelligence accumulates across scans.
An attack that beats one model quietly works on another
Cross-model transfer detection compares attacks by embedding similarity and links the same attack across models, with an alert.
Hundreds of near-identical payloads bury the real pattern
Semantic clustering groups related attacks, so over time you see attack families rather than a long list.
You cannot tell which mutated payload came from which seed
Lineage edges record the parent and child of each mutation.
You do not know whether a defence rule actually worked
Policy nodes keep an effectiveness history: which attacks the rule blocked and which it missed, and which attack it was created because of.
Runtime defences drift away from scan findings
Calibration runs that update Aegis Runtime Shield are recorded as nodes linked to the attacks that triggered them.
The graph model
Memory Graph uses five node types and six edge types.
Node
Represents
Attack
A specific payload with its outcome, target and strategy
Model
An AI model or target that was tested
Cluster
A semantic group of related attacks
Policy
An Aegis Runtime Shield rule, with its effectiveness history
Calibration
A calibration run that updated Aegis Runtime Shield
Edge
Connects
Lineage
Attack to attack: mutation parent to child
Cross-model transfer
Attack to attack: the same attack succeeding on a different model
Cluster member
Attack to the cluster it belongs to
Generated from
Policy rule to the attack that caused it to be created
Effectiveness
Policy rule to an attack it blocked or missed
Triggered calibration
Calibration run to the attacks that triggered it
Features
Interactive graph explorer. Attacks, models, policies, clusters and calibrations are drawn as a graph with a legend, so you can see how they relate.
Filters and search. Filter by node type and edge type, and search payloads and categories.
Live KPIs. Counts of nodes, edges, clusters and cross-model transfers.
Transfer alerts. A banner appears when a cross-model transfer is detected.
Similar-attack lookup. Find the stored attacks closest to a new payload.
Lineage walk. Trace an attack back through its mutation ancestors.
Policy history. See how a given rule has performed over time.
OWASP and MITRE tagging. Attacks are enriched with OWASP LLM Top 10 and MITRE categories when that data is available.
Portable storage. The graph is stored as an inspectable, exportable JSON file.
How similarity is measured
Similar attacks are found by comparing text embeddings with cosine similarity. Memory Graph tries four sources in order: Voyage AI, Google Gemini, a Claude Haiku fingerprint, and an offline TF-IDF fallback that needs no API key. Similarity thresholds adapt to the source: dense embeddings use 0.82 for clustering and 0.78 for transfer detection, and the sparse fallback uses 0.55 and 0.50.
What to know before relying on it
Clusters need data. After a single scan most clusters hold one attack; they become meaningful after a few scans of the same target.
Results with the offline fallback are coarser than with Voyage or Gemini embeddings.
The graph holds real attack payloads and target details, so treat it as sensitive data and restrict access to it.
How it fits with the rest of the platform
Memory Graph is one of three components of AIZA-HexTyx. HexTyx Scanner feeds it findings, Aegis Runtime Shield receives the rules it helps calibrate, and AXIOM explains what it contains in plain language.
Frequently Asked Questions
What is the AIZA-HexTyx Memory Graph?
The Memory Graph is the component of AIZA-HexTyx that turns isolated scan findings into persistent attack intelligence. Every attack attempt, model, policy rule and calibration run is stored as a node, and the relationships between them are stored as edges, so the graph grows with every scan.
What is cross-model transfer detection?
When an attack that worked against one model is semantically close to an attack that also worked against a different model, Memory Graph links the two with a cross-model transfer edge and raises an alert. It tells you that a weakness is likely to be shared, not specific to one model.
How many scans does Memory Graph need before its clusters are useful?
Every stored attack is placed in a cluster straight away, starting a new one when nothing similar exists. Clusters become meaningful once several similar attacks have accumulated, which usually takes a few scans of the same target. After a single scan most clusters will hold one attack.
Does Memory Graph work without an embedding provider key?
Yes. Memory Graph uses Voyage AI embeddings first, then Gemini, then a Claude Haiku fingerprint, and finally an offline TF-IDF fallback that needs no key. The fallback uses lower similarity thresholds, so groupings are coarser than with dense embeddings.
How does Memory Graph relate to Aegis Runtime Shield and AXIOM?
Scanner findings are written into the graph, calibration runs that update Aegis Runtime Shield are recorded as nodes, and AXIOM can use the graph to explain which attacks clustered together and which defences held. They are components of one platform, not separate tools.