Browse documentation

Connect AI agents

Generic MCP adapter

Use the project-neutral MCP integration pattern.

On this page

Use this adapter when an agent has Zharwing Memory MCP tools.

The daily-memory surface is complete and intentionally limited to eleven tools. Administrative daemon methods are not missing MCP capabilities; they belong to the UI and CLI administration interfaces.

Startup#

  1. Call memory.get_startup_state once for the work round with the current working directory or project id.
  2. Use its compact active/latest/recent summaries to carry forward unfinished tasks, current next steps, current blockers, and recent touched files.
  3. Do not call memory.get_latest_session when startup already supplies enough context.
  4. Call memory.start_session for a new daily/work-round session by default.
  5. Call memory.search with the task, feature, error, or file names involved.
  6. Search before requesting detail. Call memory.get_session_detail only for explicitly needed body or paginated checkpoint history.
  7. Preview or persist a context bundle only when compact state and targeted search are insufficient.

During Work#

  • Use memory.save_checkpoint after meaningful progress.
  • Supplied nextSteps and blockers replace current state; empty arrays clear it; omitted fields preserve it.
  • Treat sessions, file paths, and routine metadata in the selected project as normal AI-visible memory.
  • Use the UI or CLI for project administration, imports, graph settings, backups, Trash, or durable document editing.

Closeout#

Call memory.close_session with:

  • concrete summary
  • next steps
  • blockers, if any
  • touched files, if known

The MCP surface is intentionally limited to the daily coding-memory loop.