1. Why state is hard

• A plain LLM call is stateless: input → output, done. The moment you add a loop, tools, or multi-turn conversation, you have to answer

2. The taxonomy — five state types, by name

• 2.1 Working memory (a.k.a. scratchpad, short-term, in-task)
• 2.2 Conversational memory (a.k.a. session state, dialog history)
• 2.3 Long-term episodic memory
• 2.4 Long-term semantic memory (a.k.a. knowledge memory)
• 2.5 Procedural memory (less common, worth knowing the name)

"Agent state has five flavors: working memory for the current task trajectory, conversational memory for multi-turn dialog, long-term episodic memory for past interactions, long-term semantic memory for stable knowledge, and procedural memory for reusable skills. Each has different storage, lifespan, and retrieval patterns.”

3. The token budget problem (the central engineering challenge)

• The core tension:
• Strategy 1: Sliding window
• Strategy 2: Summarization
• Strategy 3: Selective retention (a.k.a. message filtering)
• Strategy 4: Externalized state with pointers
• Strategy 5: Hierarchical / multi-agent decomposition

4. State persistence — surviving process boundaries

The token budget problem is in-memory state management. Persistence is the orthogonal problem: what happens when the request handler dies mid-task?

• Production agents have to survive:

The pattern is checkpointing — after each meaningful step (each tool call, each turn), persist the agent's state to durable storage. On a new request, load the checkpoint and resume.