The AI memory stack, end to end

Storage stores. Security filters writes. Sgraal gates actions.

Layer 3 Execution governance

Sgraal

Preflight decision gate. Every memory-driven action returns USE_MEMORY, WARN, ASK_USER, or BLOCK. Deterministic, explainable, audit-ready.

Read the docs →
Layer 2 Write-time security

Memory filters and guards

Filter injected or poisoned content at ingest time, before it reaches storage.

ShieldCortex · A-MemGuard

Layer 1 Storage

Memory persistence and retrieval

Persist and retrieve agent memories across sessions. Vector stores, semantic indexes, episodic logs.

MemPalace · Engrama · Memori · Zep · Mem0

Why a third layer?

Storage assumes data is clean after it's written. Once a memory is persisted, every retrieval treats it as ground truth. But memory can degrade over time — timestamps age, sources lose trust, conflicting entries accumulate. Storage has no opinion about whether a memory is still safe to act on.

Write-time security filters at the gate, but attacks can evolve after ingest. A memory that was clean when written can become dangerous days later — when the context changes, when a conflicting source appears, when the action cost increases from informational to irreversible. Write-time filters don't re-evaluate.

Execution governance checks memory at the moment of action, when the cost of being wrong is highest. This is the layer that asks: is this memory safe to act on, right now, given this specific action's cost? A refund lookup and a wire transfer retrieve the same memory — but the governance decision is different because the stakes are different.

How the layers fit together

Agent
Storage retrieval
Sgraal preflight
Action

Building at Layer 2 or below?

Sgraal sits on top. We integrate with your storage and security stack — not replace it.

Talk to us