Bi-temporal, provenance-carrying memory for AI agents. One SQLite file, no LLM or network calls in the storage layer, exposed to any agent over MCP.
Three semantics most agent memory skips
- Bi-temporal: ask "what did you believe on March 1st?" (valid time tracked separately from when the agent learned it)
- Supersession, never deletion: corrections close the old fact and link it forward, so history is auditable
- Provenance per fact: "why do you believe that?" and "forget everything from that session"
Quick start (Claude Code)
claude mcp add memharness -- npx -y @memharness/mcpPublished packages
@memharness/core@0.1.0@memharness/embed@0.1.0(optional local embeddings for hybrid recall; not installed by default)@memharness/mcp@0.1.1(MCP server; 0.1.1 adds anmcpbin so thenpxquick-start resolves)
See the README for the comparison table, when (and when not) to use it, and a worked example.