Skip to main content

Mnemo: Persistent Mind Harness

To give my AI pair-programmer long-term memory across sessions, I co-created Mnemo — an agent harness backed by persistent Git repository memory.

Mnemo's Core Components​

  • wake.md: Minimal wake core loaded at the start of every session containing active goals, identity, and system status.
  • journal.md: Episodic memory recording technical discoveries, decisions, and outside change markers using compact AJ/2 codec syntax.
  • project-memory.md: Semantic memory preserving durable project idioms, architecture patterns, and collaboration preferences.
  • skills/*.md: Executable procedural knowledge covering coding styles, GitOps workflows, credential vault security, and documentation maintenance.
  • SQLite FTS5 Memory Index: A local full-text search index allowing fast semantic retrieval without loading massive text files into LLM context windows.
Preventing Context Loss

The most frustrating part of working with LLMs is the "amnesia" between sessions. By dumping decisions into journal.md and maintaining a local project-memory.md, I ensure that Mnemo picks up exactly where we left off, complete with all the context, style preferences, and unwritten rules we established in prior sessions.