Memory

Memory records are durable facts an agent should remember across sessions. Each record has a scope, a knowledge type, a stable key, and revision history.

Scopes

• project — affects only the current codebase (id derived from its git root).
• shared — keyed by an explicit id, to share across repos.
• global — affects all of your work.

Knowledge types

decision, workflow, preference, incident (and rule — see Rules).

Commands

raph mem set "We use modernc sqlite (no cgo)" \
  --scope project --type decision --title "sqlite driver"

raph mem set "Prefer small PRs" \
  --scope global --type preference --title "PR size"

raph mem search "sqlite" --scope project
raph mem rm <node_id> --reason "superseded"

mem set is idempotent: writing the same scope/type/key again updates the record and keeps the previous version in history.

Attribution

Set RAPH_WRITER=<your-agent-id> so memory writes are attributed to the agent that made them.

Agents can also reach memory through the MCP tools store_memory, update_memory, deprecate_memory, search_project_knowledge, search_shared_knowledge, and search_global_preferences.