Memory: how Iris remembers

Memory: how Iris remembers A data-flow diagram generated by Archify. 01 / Captured 02 / Recorded 03 / Distilled 04 / Stored 05 / Read back Iris sessions · terminal · headless · 01 / Captured Iris sessions terminal · headless session_store.db · 680k messages · 02 / Recorded session_store.db 680k messages Memory files · MEMORY.md + topics · 02 / Recorded Memory files MEMORY.md + topics Consolidation · Brain · 6-hourly · 03 / Distilled Consolidation Brain · 6-hourly Reflection · nightly · Sonnet · 03 / Distilled Reflection nightly · Sonnet Neo4j graph · 33k facts · 04 / Stored · Graphiti Neo4j graph 33k facts Graphiti Hindsight · pgvector · 20.5k · 04 / Stored Hindsight pgvector · 20.5k Journal + wiki · Iris-interior · git · 04 / Stored Journal + wiki Iris-interior · git iris_query · router · rerank · 05 / Read back iris_query router · rerank Session start · pack · journal · 05 / Read back Session start pack · journal every turn Stop hook lessons written by hand session notes 6-hourly day's turns nightly facts local extractor facts Gemini Flash journal append wiki pages nightly curate Cypher read-only recall read-only journal FTS read-only latest entry each session Legend primary data async batch data store data flow

Everything is kept

  • • Every turn lands in session_store.db, phone chats included
  • • MEMORY.md's 125 lines load into every session

Distilled on a timer

  • • Every 6h: notes into Graphiti and Hindsight
  • • Nightly journal reflection; Sundays compile the pack

Known gaps

  • • Phone chats are indexed but not consolidated (IRI-732)
  • • Neo4j runs without auth on all interfaces