Memory
Ódinn’s memory is a local, append-only context system with explicit provenance. It can recall accepted context during a conversation, suggest durable facts from strong user statements, and compact long sessions into summaries.
Suggested and active memory
Section titled “Suggested and active memory”The console separates suggested memories from active records. A suggestion does not become durable accepted context until you approve it.
When accepting a suggestion, choose an appropriate scope:
- Global for a preference that should apply everywhere.
- Project for decisions and context shared within one project.
- Session for context that should stay with one conversation.
Corrections supersede earlier records rather than silently rewriting history.
What memory can represent
Section titled “What memory can represent”Typical records include preferences, people, projects, decisions, procedures, artifacts, corrections, and system facts. Records can include a subject, tags, confidence, provenance, and expiry.
Keep memories concise and durable. Raw transcripts, temporary task details, and secrets are poor memory candidates.
Inspect memory from the CLI
Section titled “Inspect memory from the CLI”odinn memory search --query "weekly updates"odinn memory recall --query "how should project updates be written?"odinn memory browse --namespace userOpen a specific record:
odinn memory open --id <memory-id>Record an explicit preference:
odinn memory remember \ --kind preference \ --subject communication \ --text "Prefer concise weekly updates." \ --tags updates,styleCorrect and maintain
Section titled “Correct and maintain”odinn memory correct --target <memory-id> --text "Prefer concise biweekly updates."odinn memory compact --session <session-id>odinn memory curateMemory reads and writes remain subject to the runtime policy and audit boundary. Embedding-backed retrieval is not secretly running behind the scenes; the current system uses transparent ranked recall.
See Privacy for what can leave the machine when recalled context is sent to a cloud model.
