Mimir.

Problem.

Every new AI session started with the same explanations. Infrastructure decisions changed faster than static instructions, and useful discoveries disappeared into chat history. I wanted the next tool to pick up with evidence, decisions, and unfinished work already available.

Implementation.

1. 01

   Connected multiple AI clients through a self-hosted Basic Memory service and MCP. Plain Markdown keeps the knowledge readable outside any one assistant, while shared client configuration makes the same context available across tools.

2. 02

   Organized an entry-point index and project notes, with observations and links between decisions, implementations, and incidents. Shared guidance tells agents to consult memory before acting and record durable findings afterward.

3. 03

   Designed a hybrid ticket format: machine-readable YAML holds status, ownership, priority, and dependencies; the Markdown body preserves the reason, evidence, fix, and verification. A shared index makes unfinished work discoverable.

4. 04

   Built a scheduled private Git backup with file inventories, hashes, credential redaction, and checks that refuse to overwrite local edits. Recovery stages and reviews a snapshot before touching live notes.

Result.

Established a shared knowledge service and a consistent handoff workflow across tools. The scheduled backup published changed notes, verified their hashes, and avoided creating commits for unchanged snapshots.

What I learned.

Memory is useful when its age and authority are clear. A Git snapshot is a recovery artifact, not proof of current infrastructure state. Backups also need their own safety checks: preserve intentional Markdown formatting, redact credentials, and refuse ambiguous overwrites.