Context
What are context packs?
Context packs are reusable bundles of context you attach to a conversation deliberately. How they differ from memory and RAG, and when a curated bundle wins.
All explainers · Last reviewed:
The gap they fill
Memory is automatic: the system decides what to keep and when to reintroduce it, which is convenient and imprecise. A one-off attachment is precise and disposable — you gather the same five documents again next week.
A pack is the middle: assembled once, deliberately, and applied when you choose. Your coding standards, your product’s terminology, the constraints a piece of work has to respect.
What belongs in one
Material that is stable and that you would otherwise re-explain: house style, domain vocabulary, standing constraints, the shape of an output you always want.
What does not belong is anything that changes per question. A pack you edit every time you use it is a prompt with extra steps.
Cost and discipline
A pack is input tokens on every message it is attached to, so a large pack applied to everything is the context-window cost problem in a new shape. Several small, specific packs beat one large general one.
Because a pack is explicit, it is also reviewable — you can read exactly what is being sent, which is not true of memory that assembles itself.
Questions people ask
- How is this different from a system prompt?
- A system prompt is usually one block of instructions set once. A pack is a named bundle you attach and detach per conversation, and it can carry files and references rather than only instructions.
- Can I use several at once?
- Yes, and composing small ones is the point — a language pack plus a house-style pack rather than one bundle per project.
- Do packs replace RAG?
- No. A pack is curated by hand and always included; retrieval selects from a large corpus per question. Packs suit stable material; retrieval suits material too large to attach.
Try it rather than read about it
Context packs in ClawAI are reusable bundles you attach per conversation, so what the model receives is something you assembled rather than something inferred about you.