memory-architecturelisted

# Memory Architecture for Agents Memory is an architectural decision, not a feature. Externalizing state is hard to retrofit — make the call deliberately. ## First question: do you need memory at all? Most agentic products don't need long-term memory in v1. Conversation history within a session, plus retrieval over fixed knowledge bases, covers the majority of use cases. **You don't need long-term memory if:** - Each session is independent - Users don't expect personalization - The agent's value is in the current task, not the relationship - You can keep working state in scratchpads/files within a session **You probably need long-term memory if:** - Users return and expect the agent to remember them - Facts about the user/domain evolve and need tracking - Cross-session pattern learning is part of the value prop - The agent runs as an "ambient" assistant on long timelines Default: start without memory. Add when you have evidence of value. ## Memory types (taxonomy) Map your needs to types before choosing a vendor: | Type | What it is | Example | |---|---|---| | **Short-term / working** | Within a single session, current task state | Scratchpad, plan, intermediate results | | **Conversational history** | Full or compacted record of recent turns | Last 20 messages of a chat | | **Episodic** | Specific events, with timestamps | "Last Tuesday user asked about X" | | **Semantic** | Facts and relationships | "User's company is Acme; uses Stripe; ICP is mid-market" | | **Pro