Bank graph settings

Bank graph settings control which Hindsight banks architxt manages as contextual graphs, how often they sync, and what limits apply when importing and deploying models. You open them from the Servers page by clicking the Banks button on a server row.

This page explains each setting and destructive action in the Graph Banks dialog so you can configure a bank safely and recover from a stale or over-generated graph.

Opening the Graph Banks dialog

1. Go to Servers.
2. Find the Hindsight server that owns the bank.
3. Click Banks in the rightmost column.

The dialog lists every bank on that server. Each bank can be enabled or disabled for contextual-graph management. Only enabled banks appear in the Context Patches page selectors.

Enabling a bank

Use the checkbox on each bank row to include it in contextual-graph management. Disabling a bank removes it from the Context Patches selectors but does not delete any data.

Mode

Each enabled bank runs in one of two modes.

Mode	What it does
Manual	The bank appears in Context Patches, but nothing happens automatically. You run sync jobs by clicking Run sync job in the Context Manager.
Auto	A background sync job runs repeatedly on the configured refresh interval. Requires the sync daemon to be running.

Switching to Auto also reveals the Interval field.

Refresh interval

When a bank is in Auto mode, set how often the sync daemon should rerun the full contextual sync. Use a duration with a single unit:

• 5m — 5 minutes
• 2h — 2 hours
• 1d — 1 day
• 1w — 1 week

The dialog validates the format before saving.

Import limits

Import limits cap what the sync job brings from Hindsight into the local contextual graph.

Setting	What it controls
Top K nodes	The maximum number of Hindsight skeleton nodes to import per sync. Lower values keep the graph small; higher values capture more of the bank.
Min edge weight	The minimum weight an edge must have to be imported. Edges below this threshold are ignored.
Include patterns	Comma-separated regex patterns. Only nodes whose labels match at least one pattern are imported. Leave empty to allow everything.
Exclude patterns	Comma-separated regex patterns. Nodes whose labels match any pattern are skipped, even if they also match an include pattern.

Typical starting values are Top K nodes: 100 and Min edge weight: 0. Add exclude patterns such as temp-, noise, draft to filter out low-value nodes.

Deploy limits

Deploy limits cap what the sync job creates on top of the local graph.

Setting	What it controls
Max models per run	The maximum number of contextual mental models the sync job can deploy in one pass. This protects the bank from being flooded with generated models.
Model types	Which contextual template roles the sync job is allowed to deploy. Checked roles run; unchecked roles are skipped.

The available roles correspond to the seeded and custom contextual template roles. The four system roles are:

• Entity summary (sys_entity_summary)
• Entity capabilities (sys_entity_capabilities)
• Edge context (sys_edge_context)
• Discovery / Seed (sys_discovery_context)

Custom roles you create on Template Roles also appear here.

:::warning Auto with no restrictions If you enable Auto mode but leave both import and deploy limits empty, the dialog shows a caution warning. Auto-sync with no limits can create a large number of mental models quickly. Set at least Top K nodes and Max models per run before saving. :::

Clean and clear actions

The expanded bank row exposes several destructive operations. All of them ask for confirmation and show a toast with the result count. They do not affect source documents, entity types, or Hindsight bank settings unless explicitly stated.

Action	What it does
Dry-run clean	Simulates the Clean models action and reports how many generated mental models would be deleted. No data is changed.
Clean models	Deletes all contextual-graph mental models for this bank. Local graph nodes are marked stale but kept. Auto-sync is disabled for the bank.
Clean local graph	Removes every node and edge in the local contextual graph for this bank. Hindsight mental models are not affected. Auto-sync settings are unchanged.
Clean everything	Deletes all contextual-graph mental models and removes the entire local context graph. Auto-sync is disabled for the bank.
Clear graph & views	Removes every node and edge in the local context graph and deletes all Views for this bank. Hindsight mental models are not affected. Auto-sync settings are unchanged.
Clear all mental models	Permanently deletes all mental models from the Hindsight bank itself, including non-contextual models. This can break contextual graph or research features that rely on them.

Use the clean/clear actions when:

• A sync produced too many or the wrong models (Clean models).
• The local graph is stale and you want to rebuild it from scratch (Clean local graph or Clean everything).
• Views are stale after a major graph change (Clear graph & views).
• You need a completely blank bank state (Clear all mental models).

After cleaning or clearing, run a fresh manual or auto sync job to regenerate the parts you want.

Saving changes

Click Save Changes to write the configuration back to the server record. The save button is disabled until you actually change something. Validation runs first: any Auto bank with a malformed refresh interval blocks the save.

How bank graph settings fit into architxt

   Servers  ──►  Graph Banks dialog
                    │
                    ├── enable / disable banks
                    ├── manual or auto sync mode
                    ├── import and deploy limits
                    └── clean / clear actions
                          │
                          ▼
                   Context Patches  ◄──►  Hindsight bank

Bank graph settings are the control plane for contextual graphs. They decide which banks are managed, how aggressively they sync, and give you the tools to reset a bank's graph, views, or mental models when needed.