Memory

Extracted decision rules, logic, and agent learning from source documents — organized, searchable, and visualized as a knowledge graph.

Why Use Memory?

AI agents encode learned behavior in rules, thresholds, and decision paths. The Memory module makes this implicit knowledge explicit and auditable, letting you inspect exactly how your agents decide.

Quick Start

Get your rule extractions into AXIS in three steps:

Page Anatomy

Here is how the Memory dashboard is organized, with every major section labeled:

Summary Strip

The summary strip appears at the top of the dashboard whenever CSV data is loaded. It includes two sections:

KPI Cards

Four compact inline cards using the standard AXIS KPI pattern — icon on the left, bold value and label on the right:

• Rules — total count of extracted rules
• Risk Factors — unique risk factors across all rules
• Mitigants — total mitigants available across the rule set
• Hard Stops — unmitigated decline rules (red-tinted icon)

Charts

Two side-by-side charts provide a distribution overview:

• Rules by Action — horizontal bar chart grouped by action type (approve, decline, verify, refer, exclude). Color-coded using the AXIS chart palette.
• Rules by Product — donut chart breaking down rules by product type. The center annotation shows the total count.

Rules Tab

The default tab. Displays every extracted rule in a sortable table with multi-facet filtering and inline CRUD operations.

Filters

The filter bar sits above the table with dropdown selects for each dimension:

Filters are configurable via filter_roles in memory.yaml. The order in the YAML list determines the order in the UI. A Clear button appears when any filter is active. Changing a filter resets pagination to page 1.

Table Columns

Click any column header to sort ascending/descending:

• Rule Name — the primary identifier. Click the row chevron to expand and see the full decision path diagram.
• Action — color-coded badge (approve=green, decline=red, verify=yellow, etc.)
• Risk Factor / Risk Category — what the rule evaluates
• Product Type — which product line the rule applies to
• Mitigants — comma-separated list of available mitigating actions
• Status — ingestion status dot (green=ingested, gray=pending, red=failed)
• Actions — edit (pencil) and delete (trash) icons for inline CRUD

Pagination

The table shows 15 rules per page. Pagination controls appear at the bottom with pill-style page buttons. At most 5 page buttons are visible, centered around the current page. A "Showing X-Y of Z rules" label sits on the left.

Hard Stops Tab

Hard stops are rules matching the configured hard_stops.action_value (default: decline) with zero mitigants — absolute refusals with no workaround. This tab isolates them for focused review. Both the action value and mitigant requirement are configurable in memory.yaml.

Card anatomy:

• Red left accent border — border-l-4 border-l-red-400 on a white background (no all-red styling)
• Octagon X icon — red-tinted stop sign icon
• Title — the rule name in semibold neutral text
• Metadata line — group_by / category / product role values
• Description — the outcome_description field, line-clamped to 3 lines
• Tags — threshold type (gray pill) and compound trigger (red-50 pill) when present

Cards are paginated at 10 per page in a two-column grid.

Decision Quality Tab

This tab categorizes rules by how well they align with expected decision logic, displayed in scrollable column cards.

Three categories:

• Aligned — rules that match expected decision logic. Green check icon, count badge.
• Divergent — rules that contradict expected patterns. Amber warning icon. These need review.
• Partial — rules with incomplete alignment. Gray minus icon, displayed in a 2-column grid.

Each column has a max-h-[500px] scroll container. Below the main columns, a Soft Thresholds table lists rules with threshold_type: soft, showing rule name, risk factor, threshold value, and action.

Batches Tab

Tracks ingestion pipeline history. Each batch card shows progress and risk category coverage.

Batch card anatomy:

• Batch ID — truncated batch_id with a Layers icon
• Created date — earliest created_at timestamp in the batch
• Rule count — total rules in the batch (right-aligned)
• Progress bar — stacked bar with green (ingested), gray (pending), red (failed) segments and percentage label
• Status counts — color-coded count labels below the bar
• Risk categories — blue pill tags showing all unique risk categories in the batch

Batches are listed in a scrollable container (max-h-[600px]).

Knowledge Graph Tab

The Knowledge Graph tab connects to FalkorDB to visualize how rules, risk factors, outcomes, mitigants, and sources relate to each other in an interactive force-directed graph.

Graph Controls

• Search bar — fuzzy-search across node labels. Matching nodes highlight and the graph centers on them.
• Type filter — dropdown to filter by node type: All Types, Risk Factors, Rules, Outcomes, Mitigants, Sources.
• Zoom & pan — scroll to zoom, drag to pan the canvas.
• Click a node — opens the detail panel on the right side showing node properties, connected edges, and neighbor traversal.

Node Types

Node Detail Panel

When you click a node, a 288px-wide panel slides in from the right showing:

• Node label and type — with the type-colored icon
• Properties — all key-value pairs stored on the node
• Connected edges — list of relationships (e.g., "EVALUATES", "MITIGATES", "PRODUCES")
• Neighbor traversal — click a connected node to navigate to it

Data Format

The Memory module expects a CSV with rule extraction records. Column names are mapped to internal roles via the field_roles setting in memory.yaml. The defaults below match the standard column names.

Required Columns (default names)

These columns must exist in your CSV. Upload fails with a descriptive error if any are missing. The required set is configurable via required_roles.

id, rule_name, action, batch_id, ingestion_status

Recommended Columns

These columns are optional but enrich the dashboard significantly. If all are empty across your dataset, a warning banner appears on the Memory page.

risk_category, risk_factor, product_type, mitigants, threshold_type

Additional Optional Columns

outcome_description, threshold, compound_trigger, confidence,
decision_quality, created_at, agent_name, source, source_type,
historical_exceptions, data_fields, ingestion_error, ingested_at

Custom Column Names

If your CSV uses different column names, map them in memory.yaml:

# memory.yaml
memory:
  field_roles:
    category: compliance_area     # your column → "category" role
    group_by: risk_dimension      # your column → "group_by" role
    product: line_of_business     # your column → "product" role

Example Row

id,rule_name,action,risk_category,risk_factor,product_type,mitigants,threshold_type,outcome_description,decision_quality,batch_id,ingestion_status
R001,Max LTV Check,decline,credit,LTV Ratio,commercial,"",hard,"Decline if LTV exceeds 80%",aligned,a3f8c2d1,ingested

Key Field Values

The values below are the defaults. Each can be customized in memory.yaml under quality_values, hard_stops, and soft_threshold_value.

Configuration

The Memory module is fully configurable via backend/config/memory.yaml. No config file is needed for the default column names — every setting has a sensible default.

What You Can Customize

Quick Example

If your CSV has columns compliance_area, risk_dimension, and line_of_business instead of the defaults:

# backend/config/memory.yaml
memory:
  field_roles:
    category: compliance_area
    group_by: risk_dimension
    product: line_of_business
  labels:
    category: Compliance Area
    group_by: Risk Dimension
    product: Line of Business
  filter_roles: [action, product, category, status]

Next Steps