How should an AI agent query OneMira?
Use search before loading transcripts, request relevant segment ranges, preserve session citations, and keep MCP queries scoped for reliable agent results.
How do I connect an agent to the Local Wiki?
The Mira desktop app manages transcripts and Wiki content locally. Review extracted notes and organize durable pages according to your own project structure.
Run mira mcp list, then use mira mcp configure <client-id> for the detected client you want to connect. The local MCP server reads the same indexed OneMira data as the app and CLI.
Do not attempt to manually copy, symlink, or mount the session files into your specific project repositories.
Why: The Mira CLI maintains a local SQLite-backed index for full-text search. Use the app or CLI for writes so the readable Markdown files and index stay consistent.
How do I compress session state into project rules?
Raw conversational transcripts contain high entropy and irrelevant chatter. Relying on raw transcripts for ongoing context is highly inefficient.
Ask your agent to search first, then inspect the relevant ~/.onemira/context/<session_id>/session.md files. Review sourced decisions before writing them into a repository’s AGENTS.md or rules file.
Avoid prompting the agent to "read all meeting notes and figure out what to do next" during daily coding tasks.
Why: Reviewed project rules are shorter and easier to audit than a collection of raw transcripts. Keep source session IDs so an agent can verify a rule against its evidence.
How do I keep MCP queries bounded?
When using the local MCP server to fetch immediate context during active development, the quality of your prompt dictates the accuracy of the context retrieval.
Use explicit constraints. Example: "Use MCP to pull the 2PM meeting today regarding the Q3 marketing budget, and output the final decision."
Never use vague prompts like: "Find the meeting where we talked about the database."
Instruct your agent to use mira_search or mira_sessions (list) to get metadata and a session_id first. Then, use mira_segments with strict from_ms / to_ms or tail constraints to extract only the relevant slices.
Although modern agent harnesses can ingest an entire week of meeting transcripts in a single pass, avoid repeatedly polling the full session during a live meeting.
Why: Reading only new tail segments through mira_segments reduces transferred text and avoids repeatedly loading the full session. Actual response time depends on the client and model.
How do I inject a known session manually?
While the MCP server is powerful for autonomous agent retrieval, sometimes you just want to explicitly feed a specific meeting into your IDE (like Cursor or Windsurf) without writing a retrieval prompt.
In the OneMira Desktop App, simply right-click any session in the context list and select Copy Absolute Path or Copy Relative Path. You can then paste this path directly into your IDE (e.g., using the @ mention feature) to force-feed the exact transcript.
Why: Copying a known session reference avoids an unnecessary search step. The agent must still read and interpret the selected source correctly.