agent-memory-poisoning

Prevent memory poisoning in vector databases by implementing a deterministic memory write gate that validates entries against type schemas before persistence. This blocks attacker-controlled inputs from injecting instruction-bearing content into long-term memory stores across sessions.

Agent memory poisoning is the manipulation of long-term memory stores by attacker-controlled inputs, enabling persistent unauthorized access across sessions. It affects retrieval by allowing poisoned entries to persist in episodic buffers and vector databases, compromising agent behavior.

Enforce memory provenance by capturing source_user_id, session_id, agent_id, and timestamp for every memory write. Scoped retrieval ensures memory is retrieved with user, tenant, or role scope and labeled for provenance in prompts.

Vector database retrieval supports agent security when paired with typed memory schemas enforcing validation for preference, fact, and task_state entries. This deterministic check ensures instruction-bearing or unsafe content is prevented from persisting.

Completely delete poisoned entries by removing them from the primary store, vector index, caches, and summaries. This ensures that compromised data is purged entirely from all episodic buffers and long-term memory persistence layers.

A deterministic memory write gate is needed for episodic buffers to prevent the persistence of instruction-bearing or unsafe content. Without this gating mechanism, poisoned entries can persist across sessions and manipulate agent behavior over time.