Closedhand

How Closedhand works

Preemptive Context Layer

One live picture of your world, searched before the model reads your message.

A Preemptive Context Layer (PCL) keeps everything an assistant is connected to in one continuously updated index, a single live picture of a person’s world, and searches it automatically, without any language-model call, before the language model sees each message. The model starts with the relevant context, instead of spending tokens and tool calls deciding which apps might be worth searching. The name and this definition were first published here by Closedhand on 29 September 2026.

Why it matters

An assistant that chooses its tools can only find what it thinks to look for. Ask what to get Sam for his birthday and it may check your calendar and mail, then never open the chat where Sam said what he wanted. With a Preemptive Context Layer there is nothing to choose between. That chat is already part of the one picture, so a single search finds it alongside the calendar and mail, and the model sees connections it would never have gone looking for.

How it works in your Closedhand assistant

Everything it can read joins one index. Mail, calendars, chats and bank feeds have their own readers, and folders you choose are read file by file. Other connections are read through a common description of the information they offer: a built-in description, a published API description, or an MCP server's read-only resources and tools. A connection that can't be described this way is still used, just on demand.

It stays current on its own. Connected services are checked every 5 to 15 minutes and new items join the index as they arrive, without anyone asking. Each item is kept whole, in its own words, and a small embedding model turns it into vectors so it can be found by meaning. Files that come with anything you've connected are read too, recordings and videos written out and pictures described by a small vision language model, so they are found by what they say and show. All of it happens on the computer running Closedhand: nothing leaves it to be indexed. Apart from those picture descriptions, no language model writes into the index or decides what to fetch.

One search, in one place. Each message is searched once, by meaning and by its exact words together, in that single index. The best few matches, whichever services they came from, are placed in front of the model before it answers or picks a tool.

Detail when it’s needed. The model is shown the matching passages in their own words, which often answer the question outright. For more it reads the full original from Closedhand’s own copy on your computer, and goes to the live service only as a fallback. Tools are for acting: sending, booking or changing anything.

It belongs to you, not the model. The index lives with Closedhand on your computer, so changing your language model keeps everything it knows. Hosted models you choose see the content they work on.

What it is not

It doesn't replace tools for acting, or for live things like flight status and weather. It isn't perfect recall either: coverage depends on what each service allows, and very short or casual messages skip the search.

Fetching context before a model runs isn't new in itself. Enterprise tools, such as Microsoft 365 Copilot's semantic index, ground prompts in a company's own data. The Preemptive Context Layer names a particular form of it: for one person, across every service they connect, kept current without a language model choosing what to look up, and held by that person rather than a model provider.

Using the name

Anyone is free to use the name Preemptive Context Layer, or PCL, and to build one. If you do, please link to closedhand.com/pcl, where the name and this definition were first published. Closedhand’s own implementation is open source under Apache 2.0.

The full technical note, with how it relates to earlier work, is archived at Zenodo: doi.org/10.5281/zenodo.23040633, which always leads to its latest version.

Cite as: Closedhand, “Don’t call, recall: the Preemptive Context Layer for LLM-based personal assistants”, technical note, Zenodo, doi:10.5281/zenodo.23040633, first published 29 September 2026.