Compare Architectures

Side-by-side: use case, requirements, latency, cost, strengths, weaknesses, failure modes, example — and when to choose each.

Short-term memoryLong-term memory
Use caseWithin one session: conversation, working state, scratchpadAcross sessions: user preferences, learned facts, past outcomes
RequirementsConversation buffer plus summarization when it growsStore (KV, vector, or graph), write policy, retrieval, expiry
LatencyNone: it is already in contextRetrieval per turn plus extra tokens
CostTokens scale with conversation lengthStorage plus retrieval plus injected tokens
StrengthsPrecise, recent, no retrieval errorsContinuity and personalization
WeaknessesLost at session end; overflows long sessionsStale, contradictory, or irrelevant memories pollute context
Failure modesContext overflow, lost-in-the-middleWrong user, outdated preference, unbounded growth
ExampleThe order ID mentioned three turns ago"User prefers metric units and terse answers"
Choose this whenChoose short-term memory by default; most tasks need nothing more.Add long-term memory only for facts that must survive a session, with an explicit policy for what is written and when it expires.