Everyone says their agent "has memory"

Everyone uses the word "memory" but I feel like they all mean something different by it.

For some people it's conversation history getting stuffed back into the context window. For others it's a vector database getting queried for relevant chunks or a profile of the user that updates over time or a scratchpad the agent writes to mid-task and forgets the second the task ends.

Calling all of that "memory" hides the fact that these fail in different ways and probably need different designs entirely.

So when you say your agent "has memory," what are you actually expecting? Trying to understand your expectations and what's working / not working for you.