One call. The whole memory picture. Plain English.

What it does

Most governance APIs answer one question per call. GET /v1/insights answers all of them at once: it resolves the agent's latest memory state (from recent outcomes or the audit history), scores it, and assembles every synthesis signal into a single JSON object. The headline field is insight_summary — a one-paragraph English read of what is happening to that agent's memory, generated from the same signals.

It is a read, not a new scoring request: it does not bill a preflight, mutate state, or dispatch webhooks. Point a dashboard widget at it, or call it on a schedule to populate a health view.

The fields

A single response carries the synthesis signals an operator actually asks for:

• insight_summary — the plain-English read, composed from the signals below.
• recommended_action + omega_mem_final — the four-band decision and risk score.
• days_until_block (+ days_until_block_confidence) — the freshness-decay forecast.
• monoculture_risk_level / monoculture_risk_score — whether the agent is over-reliant on a single source family.
• single_point_of_failure_entry_id / _score — the entry whose removal would most change the decision.
• top_heal_roi_entry_id / _value — where a single repair buys the most risk reduction.
• confidence_calibration — whether the agent is calibrated or over-confident.
• memory_complexity_trend — stable, fragmenting, or forming echo chambers.
• knowledge_age_days (+ std), fleet_health_distance, cost_adjusted_decision — supporting context.

Request

curl -sS "https://api.sgraal.com/v1/insights?agent_id=support-bot-7&domain=customer_support" \
  -H "Authorization: Bearer $SGRAAL_API_KEY"

Query parameters: agent_id (required) and domain (optional, defaults to general). If there is no recent memory state for the agent, the response is { "agent_id": "...", "available": false, "reason": "no_recent_data" } — submit a preflight or an outcome first.

Response

Illustrative values shown — your response carries the same field set with your agent's own signals.

{
  "agent_id": "support-bot-7",
  "available": true,
  "insight_summary": "Agent will hit BLOCK in 5 days. Memory shows monoculture risk (HIGH). Single point of failure detected: m_014.",
  "recommended_action": "WARN",
  "omega_mem_final": 34,
  "days_until_block": 5,
  "days_until_block_confidence": 0.8,
  "monoculture_risk_level": "HIGH",
  "monoculture_risk_score": 41,
  "single_point_of_failure_entry_id": "m_014",
  "single_point_of_failure_score": 0.74,
  "top_heal_roi_entry_id": "m_014",
  "top_heal_roi_value": 31,
  "confidence_calibration": { "state": "CALIBRATED", "score": 0.62 },
  "memory_complexity_trend": "STABLE",
  "knowledge_age_days": 12,
  "knowledge_age_std_days": 9,
  "fleet_health_distance": null,
  "fleet_health_distance_available": false,
  "cost_adjusted_decision": false,
  "generated_at": "2026-05-24T09:12:04Z"
}

The decision fields carry the same semantics as the standard preflight endpoint. fleet_health_distance is null until the agent is part of a registered fleet.

Three ways teams use it

What this does not replace