SALT

SALT keeps long context small. It compresses documents and whole conversations down to the sentences that carry the most information, spreading the budget across a document's themes instead of ranking everything by one score, so the minor points that answer real questions survive. Any model, plain text out, less compute in.
-
🧂 salt
Compress a document or dataset in one shot. The engine itself, and the surface the evaluation runs on.
-
🤖 saltChat
A chat REPL where SALT is the conversation memory, so long chats and attached files stay recallable at a fixed prompt size. Each turn searches the files and the time the question names.
-
🔌 saltServe
A persistent model server chats connect to and resume against with their cache still warm.
-
🤝 Agents
Name smaller models beside the chat model and hand them work, with this conversation's memory selected for the task and nothing committed. One can answer a turn outright.
-
🔌 MCP server
salt-mcpputs compression, conversation memory and the helper models behind the Model Context Protocol, for editors and agent runtimes, and reads only when told to. -
🎛 Options
Every flag of the four commands in one line each, including the off by default switches that make long sessions better.
-
🧩 Architecture
The ideas behind the features: the keyword trie, coverage selection, the memory contract, and where each stage lives.
-
📈 Results
A 44.60 LongBench average with Llama 3.1 8B at a 20% token budget, per dataset.
-
📝 Paper
The EMNLP 2026 Main Conference paper behind all of it, with the abstract, the citation, and what changed in the repository since.
🧭 New here
- Install the environment and the four commands.
- Compress one document and read what SALT kept.
- Chat with two files attached and watch the memory block choose which one to read, and what to remember from it.
- Read why selection spreads the budget across themes instead of ranking sentences.
📜 What each version changed
The Changelog explains what every version changed in plain language.
📝 Paper
SALT is described in a Main Conference paper at EMNLP 2026, held in Budapest, and available in the ACL Anthology and as arXiv:2607.17486. The Paper page carries the abstract and the citation, and explains how the selector the paper describes, tagged v1.0.0, relates to the coverage selector current releases default to.