Memharness brings auditable, temporal memory to AI agents via a single SQLite file—no LLM calls, no network, designed for conversations and corrections that outlast your context window.

What happened: A developer released memharness, a memory system for AI agents that stores facts with two time dimensions—when something became true in the world (valid_from/valid_to) and when the agent learned it (tx_at)—plus provenance tracking for each fact. The system runs as a library or MCP server (a standard interface for connecting tools to AI clients) and uses plain SQLite, with no LLM or network calls in the storage layer.

Why it matters: Agent memory often grows too large for context windows (the text limit in each conversation), and bag-of-string storage cannot answer audit questions like "what did you believe before I corrected you?" or "why do you believe that?". Memharness answers those by keeping old facts as history (never deleting) and tagging each fact with its source, making it possible to revert changes, time-travel queries, and selectively forget information by source—useful for compliance (GDPR-shaped deletion) and explainability.

What to watch: The system is positioned as a storage layer, not an automatic fact extractor; you decide what to remember. It is available now as a library (@memharness/core) and MCP server (@memharness/mcp), with optional semantic search via a local embedding model (BGE-small, ~130MB). The default database lives at ~/.memharness/memory.db with no signup or API key required.