Moyu AI Agent Secure Memory Toolkit

Problem

Persistent memory in AI agents creates three concrete risks: injected or poisoned writes, PII such as phone numbers, IDs, and API keys left in local files, and loss of key decisions when context limits force truncation. Many memory systems depend on external vector stores, LLM APIs, or cloud tenancy, which makes offline and zero-infrastructure deployment awkward.

How it works

Moyu treats the memory layer as a stack of defense, retrieval, and lifecycle stages. Before writes, content passes through 613 regex checks and LLM review; suspected injection, PII, or abnormal write bursts are blocked, masked, or rolled back. Retrieval combines semantic, BM25, recency, entity, and staged-content signals, then uses search feedback to tune weights. After storage, memories are compressed, forgotten, and merged by heat, scene, and semantic review, while original text is retained for audit. It favors local execution: ONNX embeddings, SQLite FTS5, and Markdown files carry the core state; LLM features are on by default but degrade to rule-based paths without an API key.

Limits

The defense chain is layered deterrence, not a silver bullet. It covers known injection patterns, accidental corruption, and common PII reasonably well, but targeted bypasses or metaphor-style semantic injection can still succeed. Strict compliance, cross-tenant isolation, high availability, or advanced knowledge-graph requirements should be handled with an organization-level gateway, encryption, and audit strategy rather than relying only on a local memory toolkit.