Skip to content
amirhouiehPublic

About

A self-evolving memory layer for LLMs.

Topics

Resources

Stars

33 stars

Watchers

0 watching

Forks

Repository files navigation

adapt

adapt is a TypeScript memory layer designed around attention, self-evolution and proactivity, optimized for small local and on-edge models (1–2B).

adapt takes a different approach to memory. Most memory layers are recall-based: they store data and retrieve it when asked. adapt is attention-based: it tracks what you tell it to, and decides which incoming data deserves its effort. This lets adapt keep a small, current memory instead of a growing archive, notice data that fits nothing it tracks, and reshape what it tracks as the data changes. Each of these steps is narrow enough for a small local model.

adapt is where attention, self-evolution and proactivity overlap.

Under the hood, adapt splits memory into neurons: small units, each attending to one topic and keeping one plain-text memory of it, called its understanding. A brain is the set of neurons that serves one goal, which you give as a prompt. inject() sends new text to every neuron; each extracts what fits its topic and rewrites its understanding. Text no neuron keeps is recorded as a gap, and gaps drive evolution, which creates, merges or removes neurons. ask() sends a question to the neurons whose topics match and merges their answers.

Install

npm i adapt-memory

Quickstart

Load lfm3-1.7b in LM Studio with a 32768-token context and start its server on port 1234. Then:

import { Brain } from 'adapt-memory'
import { SQLiteBrainStore } from 'adapt-memory/sqlite' // npm i better-sqlite3

const brain = await Brain.create({
  prompt: 'Track my coding preferences and food habits.',
  store: new SQLiteBrainStore('./brain.db'),
})
await brain.inject(['I always use strict TypeScript.', 'Dinner: vegetarian ramen.'])
const answer = await brain.ask('What do I prefer in code?')
console.log(answer.insight, answer.gaps)
await brain.dispose()
// later: const brain = await Brain.restore('./brain.db')

The code needs ESM and top-level await. The quickstart walks through each step.

Status

MIT. Version 0.2.0 (changelog). Earlier versions were published as @unbody-io/adapt; their stores are not supported. Shipped features, next steps and limits are on the concepts page.

Links

About

A self-evolving memory layer for LLMs.

Topics

Resources

Stars

33 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages