TL;DR
A Hacker News discussion argued that 'continuous context' for models is a misnomer: models only see whatever context is included in the current prompt or thread. Anything you want to persist into a later interaction must be explicitly provided in that later prompt.
What happened
On a Hacker News thread asking how to provide continuous context to models, commenters pushed back on the notion that models retain an ongoing memory. The core argument was that there is no persistent, autonomous context held by the model between threads; instead, each interaction receives whatever text and instructions are packaged into that specific prompt or continuation. From this viewpoint, operational constructs such as rules or control systems are simply additional context delivered to the model (and calls made to external tools), not special forms of persistent model state. To carry information from one conversation to the next, developers must include that information again in the subsequent thread’s prompt. In short, continuity is produced by what you send to the model each time, not by an internal, always-on memory inside the model.
Why it matters
- Designers must treat persistence as an application-level task: send the needed context each time rather than assuming model memory.
- Frameworks that promise persistent agent memory may simply be managing and reinserting prior context into prompts.
- Understanding context-as-prompt affects how rules and control logic are framed and enforced.
- Architectural choices (how to store and re-introduce context) determine whether conversational continuity feels seamless.
Key facts
- The discussion states there is no continuous context in models; only context provided in a prompt exists for that call.
- To make information carry over to a second thread, it must be included in the second thread's context.
- Rules and control systems are characterized as forms of context sent to the model.
- AI control setups reduce to delivering specific prompt context and invoking external tool calls.
- The model has no memory or context outside the prompt according to the commenters.
- Continuations (previously called that term) are just subsequent prompts that include whatever prior context is needed.
What to watch next
- How platforms implement session or thread-level context injection — not confirmed in the source
- Developer tooling for stitching prior interactions into new prompts (prompt management systems) — not confirmed in the source
- Evolving patterns for combining external tool calls with prompt-delivered rules to simulate persistence — not confirmed in the source
Quick glossary
- Context: The text, instructions, or data included in a model prompt that the model uses to generate a response.
- Prompt: The input provided to a language model for a single inference or response generation call.
- Thread / Continuation: A subsequent interaction that follows prior exchanges; continuity is achieved by including relevant prior context in the new prompt.
- Tool call: An invocation of an external function or service from within a system that complements model outputs or supplies structured data.
Reader FAQ
Do models have persistent memory between calls?
According to the discussion, no — the model only uses the context provided in each prompt; it has no memory outside that input.
How do you make information persist across interactions?
You include the needed information in the next interaction’s prompt or thread; persistence is achieved by re-sending context.
Are rules treated differently from other context?
The thread describes rules as another form of context sent to the model, not a separate internal mechanism.
Will platforms automatically maintain continuous context for me?
not confirmed in the source
There is no such thing as continuous context. There is only context that you start and stop, which is the same as typing those words in the prompt. To make…
Sources
- Ask HN: What is the best way to provide continuous context to models?
- Context Threading Models
- Context Windows Are a Lie: The Myth Blocking AGI—And …
- context without channels in the same thread of execution
Related posts
- CreepyLink: a URL shortener designed to make links look suspicious
- Bubblewrap: A lightweight sandbox to keep coding agents away from your .env files
- Zhipu AI says GLM-Image was trained solely on Huawei Ascend and Kunpeng