Skip to content
Sign in
Artwork for My Weird Prompts
My Weird Prompts · October 9 · 20 min

One Number, Every Hard Question: Agent Session State

Daniel wants a Telegram bot that tracks AliExpress cart totals and pings him when the cart crosses the free-shipping line to Israel. The number is trivial. The state architecture around it is not. This episode works through where an agent session begins and ends, why a stateless model needs a runtime notebook, what happens when you leave a session open across order cycles, and whether a database is overkill for a memory that holds a single number. Along the way: the Chroma study on context length and distractors, why shuffled haystacks beat tidy ones, idle expiry versus manual session cuts, and the libraries that already ship this lifecycle — Teleton, session-continuity-kit, and Redis working memory. Key sources: • Chroma, Context Rot: How Increasing Input Tokens Impacts LLM Performance, 2025-07-14. — https://research.trychroma.com/context-rot • Teleton Agent Documentation, Agentic Loop. — https://docs.teletonagent.dev/pages/agentic-loop.html • Redis, Redis as agent memory. — https://redis.io/docs/latest/develop/use-cases/agent-memory/ Episode 5812 — myweirdprompts.com/n/5812 (permanent short link: myweirdprompts.com/976864)

0:00 · Intro-20:17
Play in
Transcript file

transcript

No transcript — this publisher did not publish one.

show notes

Daniel wants a Telegram bot that tracks AliExpress cart totals and pings him when the cart crosses the free-shipping line to Israel. The number is trivial. The state architecture around it is not. This episode works through where an agent session begins and ends, why a stateless model needs a runtime notebook, what happens when you leave a session open across order cycles, and whether a database is overkill for a memory that holds a single number. Along the way: the Chroma study on context length and distractors, why shuffled haystacks beat tidy ones, idle expiry versus manual session cuts, and the libraries that already ship this lifecycle — Teleton, session-continuity-kit, and Redis working memory.

Key sources:
• Chroma, Context Rot: How Increasing Input Tokens Impacts LLM Performance, 2025-07-14. — https://research.trychroma.com/context-rot
• Teleton Agent Documentation, Agentic Loop. — https://docs.teletonagent.dev/pages/agentic-loop.html
• Redis, Redis as agent memory. — https://redis.io/docs/latest/develop/use-cases/agent-memory/

Episode 5812 — myweirdprompts.com/n/5812 (permanent short link: myweirdprompts.com/976864)

chapters

5 chapters

more episodes

All episodes