epic(memory): memory security hardening

**Context:** ROADMAP.md → Memory security (entire section)
**Related:** #27 (review feedback memory), #26 (closed input hardening)

---

## Component

Agent (Python runtime)

## Describe the feature

Epic tracking **memory security** capabilities for AgentCore Memory: trust-aware retrieval, decay, anomaly detection, quarantine, write-ahead validation, review quorum, backup, extraction replay, and optional knowledge graph (tier 4).

## Use case

Memory is a learning surface and an attack surface. Poisoned or stale memories degrade agent quality or enable prompt injection persistence. Operators need detect-and-respond controls without blocking the fail-open execution path for normal tasks.

## Proposed solution

Child issues (file separately):

| Child draft | Capability |
|-------------|------------|
| [memory-trust-aware-retrieval.md](./memory-trust-aware-retrieval.md) | Trust-aware retrieval |
| [memory-temporal-decay.md](./memory-temporal-decay.md) | Temporal decay |
| [memory-anomaly-detection.md](./memory-anomaly-detection.md) | Anomaly detection |
| [memory-quarantine-and-rollback.md](./memory-quarantine-and-rollback.md) | Quarantine and rollback |
| [memory-write-ahead-validation.md](./memory-write-ahead-validation.md) | Write-ahead validation |
| [memory-review-feedback-quorum.md](./memory-review-feedback-quorum.md) | Review feedback quorum |
| [memory-backup-to-s3.md](./memory-backup-to-s3.md) | Memory backup to S3 |
| [memory-extraction-replay.md](./memory-extraction-replay.md) | Memory extraction replay |
| [memory-structured-knowledge-graph.md](./memory-structured-knowledge-graph.md) | Structured knowledge graph (tier 4) |

Ship incrementally; memory writes remain fail-open for task execution unless operator enables enforce mode per repo.

## Other information

- Design context: `docs/design/MEMORY.md`, `docs/design/SECURITY.md`.
- #27 implements acquisition; this epic implements **security and lifecycle**.

- [ ] This might be a breaking change

Commencez par ROADMAP.md et la section sur la sécurité de la mémoire, puis lisez docs/design/MEMORY.md et docs/design/SECURITY.md. Examinez les brouillons enfants listés pour comprendre les capacités distinctes et leur portée incrémentale. Le travail est considéré comme terminé lorsque l’epic est découpé en issues enfants implémentables, tout en conservant le comportement par défaut fail-open et le mode enforce optionnel.

Rédigé par le modèle d'indexation à partir du texte de l'issue.