LEARN — how Harvey accumulates and retrieves knowledge
Harvey stores knowledge in three independent silos that are unified at retrieval time. Understanding which silo to use for which kind of content is the key to getting consistently good results.
Have a text file or document? → /rag ingest
Silo What belongs here How to add How it arrives ───────────────
───────────────────────── ─────────────── ───────────────────── RAG
store Reference docs, API specs, /rag ingest Per-prompt via code
examples, PDF papers
Memory store Patterns observed during /memory mine Session-start via sessions: what worked, (interactive) /memory recall what the model got wrong, auto-mines on user preferences session exit
Knowledge base Research notes, named /kb observe On-demand via experiments, cross-project /kb project /memory recall concepts and hypotheses /kb concept
Retrieval from all three silos is unified:
/memory recall
/profile <list|show|edit|use|rename> [args…] — alias for /memory profile (manage workspace profile) /profile list — list active and archived profiles /profile show — print full content of the active profile /profile edit — open active profile in $EDITOR /profile use name — switch profile: saves handoff, archives old profile, selects new template, resets history /profile rename NAME — rename the workspace in the active profile
/status — shows active memories, unmined sessions, RAG chunk count /hint — prints actionable suggestions for improving results /memory status — detailed memory store stats and budget advice /rag status — RAG store details: active store, chunk count, on/off
Ingest a PDF reference before starting a coding session:
/rag ingest Reference/papers/oberon2.pdf /rag on
Mine learnings from last session before starting the next:
/memory mine
Record an observation about a running experiment:
/kb observe “Qwen3.5 handles nil pointer chains correctly after explicit cast”
Check what would improve the current session:
/hint
/help rag — full RAG command reference /help memory — full memory command reference /help kb — full knowledge base reference