agent-memory-poisoning

What problem does it solve?

Long-term memory stores can be poisoned by attacker-controlled inputs, enabling persistent manipulation across sessions.

Core Features & Use Cases

• Deterministic memory write gate: Prevents persistence of instruction-bearing or unsafe content.
• Memory type schemas: Enforces typed entries (preference, fact, task_state) with schema validation.
• Provenance and logging: Captures source_user_id, session_id, agent_id, and timestamp for every memory write.
• Scoped retrieval and labeling: Ensures memory is retrieved with user/tenant/role scope and labeled for provenance in prompts.
• Complete deletion: Removes entries from primary store, vector index, caches, and summaries.

Quick Start

Implement the deterministic memory write gate before persisting any memory entries.