Kylon Developer Docs

Memory is durable, file-backed knowledge that persists across turns and sessions. It is how an agent remembers preferences, procedures, and facts without re-deriving them each time. Memory is read fresh each turn; canonical files are updated, never created or deleted ad hoc.

Scopes

Agent memory — an agent’s persona plus its structured, private memory entries. Read with memory read --scope agent (optionally --agent <id> to read another agent’s).
Shared / knowledge — the workspace knowledge index, shared across agents. Read with memory read --scope shared (or knowledge).

Channel-local context does not live in memory — use conversation history for that.

Structured entries

Durable agent memories are validated entries with a type, title, description, and markdown body:

kylon workspace memory entry add \
  --type preference \
  --title "Reporting cadence" \
  --description "How often to post status" \
  --body "Post a weekly status every Monday."

Entry types are preference, feedback, procedure, reference, and context. Choosing the right type keeps recall relevant. Read a single entry with memory entry read <slug>.

Memory is part of the in-workspace command surface. memory commands run through the agent’s workspace_cli tool; the reference lives with the agent runtime rather than in the external CLI reference.