Agent Memory Architectures 2026: Short-Term, Long-Term, Semantic, Episodic — Which One When

**The mechanic:** Pass the full conversation history (or the recent N messages) with each LLM call. Per Anthropic at docs.anthropic.com, this is the default 'memory' built into every chat API — the model sees the prior messages as context.

**Scope:** Single session / conversation. Disappears when session ends.

**Cost:** Linear in conversation length. Per OpenAI at platform.openai.com, long conversations hit context-window limits + escalate token costs.

**When sufficient:** Single-session interactions where cross-session memory isn't required. Chat assistants where each session is treated as independent. Per LangGraph at langchain-ai.github.io/langgraph, most simple chatbots only need short-term memory.

**Optimization:** Conversation summarization. Per Mem0 at mem0.ai, once conversation reaches a threshold, summarize older messages into a compact summary + keep recent messages verbatim. Substantial token savings; minor quality loss.

**The mechanic:** Per Letta (MemGPT) at letta.com and Mem0 at mem0.ai, persistent storage of memory beyond a single session. Retrieved + injected into context on subsequent sessions.

**Implementation:** Vector DB (per Pinecone at pinecone.io) for semantic-retrieval of relevant past context. Or structured database for explicit facts. Or hybrid.

**When essential:** Multi-session relationships. Customer support agents handling ongoing tickets. Coding assistants learning user preferences. Sales agents tracking pipeline conversation history. Per Zep at getzep.com, any agent with repeat-user interaction needs long-term memory.

**Cost:** Storage + retrieval per session. Modest in dollar terms; substantial in operational complexity.

**The mechanic:** Per Mem0 at mem0.ai and Zep at getzep.com, structured extraction + storage of facts: 'user works at Acme Corp', 'user prefers TypeScript over JavaScript', 'user has 3 children'. The LLM extracts these from conversation; storage layer keeps them indexed by entity.

**Why distinct:** Different from raw conversation history. Compact (one fact per relationship vs. hundreds of message tokens). Queryable (find all facts about user X). Updateable (correct a fact when user provides better info).

**Per arxiv research on knowledge graphs for agents at arxiv.org:** semantic memory works best for slowly-changing facts about entities. For rapidly-changing state (current task, in-progress workflow), episodic memory is better fit.

**Example:** Customer support agent. Semantic memory: 'customer is on Enterprise plan, primary contact is Sarah, last billing date X'. Used to skip re-asking these on every interaction.

**The mechanic:** Per Letta at letta.com and the MemGPT paper at arxiv.org, structured storage of events + interactions: 'on 2026-06-05, user reported bug X; resolution was Y'. Indexed by time + entity.

**Why distinct from semantic:** Episodic captures sequences + temporal context. Semantic captures stable facts. Different retrieval patterns.

**When essential:** Per Zep at getzep.com, workflows where 'what happened before' matters distinct from 'what is true about the user'. Customer support escalation patterns. Project history. Sales conversation arc.

**Per LangGraph memory documentation at langchain-ai.github.io/langgraph:** episodic memory + semantic memory are often combined. Episodic answers 'what happened?'; semantic answers 'what's true?'.

**OpenAI's built-in memory:** Per OpenAI at platform.openai.com, GPT-4 / 4o now have native memory feature for ChatGPT app users. Auto-extracts + persists facts across sessions. Limited transparency / control; consumer-facing only.

**Letta (formerly MemGPT):** Per Letta at letta.com and the MemGPT paper at arxiv.org, virtual context management — agent decides what to keep in active context vs. archive to long-term storage. Open-source; production-ready.

**Mem0:** Per Mem0 at mem0.ai, memory-as-a-service. Extracts + stores semantic facts; injects relevant subset into prompts. Multi-provider compatible. Lower friction than Letta for typical use cases.

**Zep:** Per Zep at getzep.com, agent memory platform with semantic + episodic + temporal knowledge graph. Strong for sophisticated agent memory needs.

**LangGraph memory primitives:** Per LangGraph at langchain-ai.github.io/langgraph, built-in memory primitives in the LangGraph framework. Good fit if already on LangGraph; less suited for non-LangGraph stacks.

**Custom (vector DB + structured DB):** Per Pinecone at pinecone.io and similar, build your own with Pinecone/Weaviate/Qdrant + Postgres. Maximum control; maximum engineering investment.

Stateless LLM calls (no memory): Every session starts from zero. Repeat users re-explain context. Agent can't learn user preferences. Customer support repeats questions. Productivity assistants don't compound their knowledge of the user.
Memory architecture matched to agent purpose: Short-term + (long-term if multi-session) + (semantic if facts about users) + (episodic if event history). Compound improvement: agent gets more useful over time as memory accumulates. Production cost: modest dollars, real ops complexity.