Documentation

ZMem is local-first memory for AI agents with receipts for what influenced every action. The product goal is simple: agents should remember across sessions without turning memory into an unverifiable black box.

Use these docs when you want to install ZMem, connect an agent, inspect memory status, hand work to another environment, or run benchmark evidence without overstating what the numbers mean.

The Product Contract

• Local memory is the source of truth by default.
• Memory that can influence an action should be reviewable, governable, and revocable.
• Every important action should be explainable through receipts, bundles, snapshots, hashes, and optional Treeship proofs.
• Benchmarks should be reproducible from local artifacts before they become public claims.

Where To Start

• Start with Install for a clean local setup.
• Use Agents to connect Codex, Claude Code, Cursor, OpenClaw, Hermes, or generic MCP clients.
• Read Builders if you are adding ZMem to a tool, agent, workflow, or product surface.
• Read Receipts before sharing proof artifacts.
• Run Benchmarks when you need retrieval quality, latency, token, and proof evidence.

Public Docs Map

• Users: install, connect an agent, review memory, inspect receipts, and restore handoffs.
• Builders: embed ZMem through the CLI, MCP server, local database, receipts, and handoff artifacts.
• Teams: benchmark retrieval behavior and decide what evidence is safe to publish.