229 stories
Short-term memory in an LLM is whatever sits in the context window on the current call: the system prompt, the recent conversation, tool results and any retrieved text, all re-sent by the application on every request and gone when the request ends.
GPT-4o has a 128,000-token context window and can write at most 16,384 tokens in a single reply, with the reply counted inside the same 128,000.
A long conversation costs far more than its number of turns suggests, because every request carries the system prompt and the entire history again, so total input grows with the square of the number of turns.
The Claude API costs between $1 and $10 per million input tokens and between $5 and $50 per million output tokens on Anthropic's current models.
Nine Claude Skills we run in production, plus one bonus, with the install command for each and what it saves you. Updated September 2026.
Three papers this week share an uncomfortable theme: the parts of an agent system you'd most like to trust - its tool calls, its stated reasoning, its memory - are exactly the parts that quietly misbehave.
Six things we learned adding Jev, a model that returns typed decisions instead of text, as a reranker in Hindsight 0.10.1. Including two designs that failed and the numbers that killed them.
Managed Deep Agents is the simplest way to build, deploy, and run agents in production. The 0.8 release adds support for user-owned credentials, user-level memory, HTTP channels, file transfer in Slack and a pre-built tool for web search.
We spent a year teaching an AI agent to remember. The headline feature returned null on 31 of 31 calls. What failed, why, and the two things that worked.
When you migrate a memory history, its duplicate conclusions come with it. The fix is to import the source material and let your bank rebuild the beliefs itself.
Spotify wrote its routing rules into CLAUDE.md and Claude could ignore them; a hook that blocks whole-file reads over 350 lines could not be ignored. 65 steps, templates and file formats, each with its source, sorted by what you are trying to build.
A practical definition of AI-native, the architecture behind it, real examples, and a 10-point test for separating genuine AI-native products from AI add-ons.
Clone a memory bank in one call, export and import it with an explicit scope, or rename it in place. What travels, what gets rebuilt, and where it will bite you.
The best chatbot memory layer plugin captures durable user information, stores it with the right scope, retrieves only the relevant part of it, and updates or deletes it reliably.
Learn how to use Engram effectively, from writing topic descriptions that control extraction to choosing bounded topics and scopes, selecting a retrieval mode, and placing memories in your prompt without hurting prompt-cache efficiency.
The AI story of September 14–20 was not one spectacular model launch. It was the machinery around agents becoming visible: people reading private chats, models writing instructions into their own memory, plugins updating beneath users, and regulators asking how to stop a system after release.
Six releases in eight weeks, from 0.8.5 to 0.10.0. The biggest things Hindsight learned to do this summer, grouped by what they let you build.
Learn how to engineer the instructions, state, retrieval, memory, tools, and evaluation that reliable AI agents need at runtime.
Showing 126 of 229 stories
Load more storiesGet the top AI news and tools straight to your inbox every week.
Weekly · AI news
Join 15,000+ AI enthusiasts
