Keep less. Decide the same.

What it does

Sgraal scores your full memory state to get a baseline decision. Then, for each entry, it tests whether removing it changes that decision. Entries whose removal leaves the decision unchanged are removable; the rest form the minimum_subset — the decision-preserving core.

The result is not a guess: each removable entry was verified by an actual re-score. You get back the essential subset, the removable list with a reason per entry, and an estimated_savings summary.

Why it matters · 1 — Data-minimization

GDPR Article 5(1)(c) requires personal data to be “adequate, relevant and limited to what is necessary.” For an AI agent, “necessary” is hard to argue in the abstract — but MVMem makes it concrete and auditable: the decision-preserving subset is the necessary set for the action being governed.

Drop the rest, and you reduce both your data-retention surface and the personal data exposed to downstream models — with a re-scored record showing the kept set still produces the same decision. A DPO can point at that record.

Why it matters · 2 — Token-cost ROI

Every entry you carry into an agent's context window is tokens you pay for on every call. If a chunk of memory does not change the decision, it is paying rent for nothing.

A simple worked example: an agent carries ten retrieved memories into each prompt; MVMem finds four of them are removable without changing the decision. Trimming to the decision-preserving subset cuts the memory portion of the prompt accordingly — on every subsequent call, indefinitely. The savings compound with call volume; the exact figure is your own token economics, computed from the estimated_savings.entries_removed the endpoint returns.

Request

curl -sS https://api.sgraal.com/v1/mvmem \
  -H "Authorization: Bearer $SGRAAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d @request.json

request.json carries memory_state (the entries to minimise), plus optional action_type, domain, and agent_id — the same memory-entry shape as /v1/preflight.

Response

Two of ten entries are removable without changing the decision. Illustrative values shown.

{
  "original_decision": "USE_MEMORY",
  "original_omega": 16,
  "minimum_subset": ["m_002", "m_005", "m_009"],
  "removable_entries": [
    { "id": "m_001", "reason": "removal does not change decision" },
    { "id": "m_003", "reason": "removal does not change decision" }
  ],
  "estimated_savings": { "entries_removed": 2, "percent": 40 }
}

For large memory stores the endpoint samples before testing, and marks the unsampled entries accordingly in removable_entries — so the result stays fast without claiming to have tested entries it did not.

Pairs with the MVMem Certificate

When you need to prove the minimisation — for an auditor, a DPO, or a counterparty — the MVMem Certificate issues a cryptographically signed record that the kept subset preserves the decision. The endpoint here gives you the subset; the certificate makes it externally checkable.

What this does not replace