Mnemosyne Local AI Memory System

Problem Context

When adding long-term memory to an AI agent, common issues include high context cost, weak keyword retrieval, unauditable records, and services becoming unavailable when optional vector components are missing. Mnemosyne treats memory storage as a local-first, pure Python standard-library capability, reducing hard dependencies and cloud lock-in while recalling only high-value content under a limited token budget.

How It Works and Boundaries

The core engine centers on retain for writes, recall for retrieval, graph-query for graph lookup, and verify-integrity for audit. Memory is organized into hot, warm, and cold tiers, where low-value records migrate to the cold tier instead of being deleted. Writes enter a SHA-256 hash-chain ledger that can detect and locate tampered records, while entity relationships are stored as subject-predicate-object triples and support path-expanded retrieval. Chinese workflows use bigram tokenization, FTS5, and a synonym dictionary, with pre-write checks for credentials, invisible Unicode, and HTML injection. Optional plugins such as cryptography, numpy_vector, and reranker degrade gracefully when dependencies are missing, keeping the main flow intact. Note: the 90%+ token savings figure is a reproducible project-reported metric, not a guarantee that every end-to-end LLM workload will reduce tokens by the same ratio; vector semantic search still requires numpy and optional models, and is not part of the zero-dependency core.