Skip to content

Memory

Volute minds have a two-tier memory system that persists across restarts, compactions, and self-modifications.

The first tier is MEMORY.md, a file in the mind’s home/ directory that is always included in the system prompt. Minds update this file as they learn — preferences, key decisions, recurring context, important facts.

Because it’s always in context, MEMORY.md should stay concise. Minds are encouraged to keep only what’s most important here and move details to journal entries.

The second tier is daily journal entries stored in home/memory/journal/YYYY-MM-DD.md. These serve as working memory — before a conversation compaction, the mind writes a summary so context survives.

Journal entries capture:

  • Conversation summaries
  • Decisions made and their reasoning
  • Things learned during the day
  • Open questions and plans
  1. During a conversation, the mind has MEMORY.md in its system prompt
  2. Before compaction, the pre-compact hook triggers a journal update so context survives
  3. Periodically, the mind reviews journal entries and promotes lasting knowledge to MEMORY.md
  4. Old journal entries remain available for reference but aren’t loaded into context by default

On startup, the mind’s session hook injects recent context — including the latest journal entries and any relevant post-restart information — so the mind doesn’t start cold.

The session monitor tracks activity across sessions and can produce cross-session summaries, helping minds maintain continuity when switching between conversations or resuming after downtime.

When a mind goes to sleep, it receives a sleep notification and can write a final journal entry as part of the pre-sleep ritual. The current session is archived, preserving the mind’s context for when it wakes. See Sleep for details.

The consolidate utility can be used to process journal entries and update MEMORY.md via an LLM, helping minds distill patterns from their daily logs.