Skip to content

agentailor/slimcontext

Repository files navigation

slimcontext

Lightweight, model-agnostic chat history compression utilities for AI assistants. Bring Your Own Model (BYOM) and use simple strategies to keep conversations concise while preserving context.

CI

Supported Strategies

Trimming

Simple token-based compression that removes the oldest messages when your conversation exceeds the token threshold. Always preserves system messages and the most recent messages to maintain context continuity.

Trimming Strategy

Summarization

AI-powered compression that uses your own chat model to create concise summaries of older conversation segments. The summary is injected as a system message, preserving the conversation flow while drastically reducing token usage.

Summarization Strategy

Examples

Features

  • Trim strategy: token-aware trimming based on your model's max tokens and a threshold.
  • Summarize strategy: token-aware summarization of older messages using your own chat model.
  • Framework agnostic: plug in any model wrapper implementing a minimal invoke() interface.
  • Optional LangChain adapter with a one-call helper for compressing histories.

Installation

npm install slimcontext

Migration

Upgrading from an earlier version? See the Migration notes in the changelog:

  • CHANGELOG: ./CHANGELOG.md#migration

Core Concepts

Provide a model that implements:

interface SlimContextMessage {
  role: 'system' | 'user' | 'assistant' | 'tool' | 'human';
  content: string;
}
interface SlimContextModelResponse {
  content: string;
}
interface SlimContextChatModel {
  invoke(messages: SlimContextMessage[]): Promise<SlimContextModelResponse>;
}

slimcontext handles message arrays shaped as:

interface SlimContextMessage {
  role: 'system' | 'user' | 'assistant' | 'tool' | 'human';
  content: string;
}

Usage

TrimCompressor

import { TrimCompressor, SlimContextMessage } from 'slimcontext';

// Configure token-aware trimming
const compressor = new TrimCompressor({
  // Optional: defaults shown
  maxModelTokens: 8192, // your model's context window
  thresholdPercent: 0.7, // start trimming after 70% of maxModelTokens
  minRecentMessages: 2, // always keep at least last 2 messages
  // Optional estimator; default is a len/4 heuristic
  // estimateTokens: (m) => yourCustomTokenCounter(m),
});

let history: SlimContextMessage[] = [
  { role: 'system', content: 'You are a helpful assistant.' },
  // ... conversation grows
];

history = await compressor.compress(history);

SummarizeCompressor

import {
  SummarizeCompressor,
  SlimContextMessage,
  SlimContextChatModel,
  SlimContextModelResponse,
} from 'slimcontext';

class MyModel implements SlimContextChatModel {
  async invoke(messages: SlimContextMessage[]): Promise<SlimContextModelResponse> {
    // Call out to your LLM provider (OpenAI, Anthropic, etc.)
    const userContent = messages.find((m) => m.role === 'user')?.content || '';
    return { content: 'Summary: ' + userContent.slice(0, 100) };
  }
}

const model = new MyModel();
const compressor = new SummarizeCompressor({
  model,
  // Optional: defaults shown
  maxModelTokens: 8192,
  thresholdPercent: 0.7, // summarize once total tokens exceed 70%
  minRecentMessages: 4, // keep at least last 4 messages verbatim
  // estimateTokens: (m) => yourCustomTokenCounter(m),
  // prompt: '...custom summarization instructions...'
});

let history: SlimContextMessage[] = [
  { role: 'system', content: 'You are a helpful assistant.' },
  // ... conversation grows
];

history = await compressor.compress(history);

Notes about summarization behavior

  • When the estimated total tokens exceed the threshold, the oldest portion (excluding a leading system message) is summarized into a single system message inserted before the recent tail.
  • The most recent minRecentMessages are always preserved verbatim.

Strategy Combination Example

You can chain strategies depending on token thresholds or other heuristics.

Example Integration

Adapters

LangChain

If you already use LangChain chat models, you can use the built-in adapter. It’s exported in two ways:

  • Namespaced: import { langchain } from 'slimcontext'
  • Direct path: import * as langchain from 'slimcontext/adapters/langchain'

Common helpers:

  • compressLangChainHistory(history, options) – one-call compression for LangChain BaseMessage[].
  • toSlimModel(llm) – wrap a LangChain BaseChatModel for SummarizeCompressor.

Example (one-call history compression):

import { AIMessage, HumanMessage, SystemMessage } from '@langchain/core/messages';
import { ChatOpenAI } from '@langchain/openai';
import { langchain } from 'slimcontext';

const lc = new ChatOpenAI({ model: 'gpt-5-mini', temperature: 0 });

const history = [
  new SystemMessage('You are helpful.'),
  new HumanMessage('Please summarize the discussion so far.'),
  new AIMessage('Certainly!'),
  // ...more messages
];

const compact = await langchain.compressLangChainHistory(history, {
  strategy: 'summarize',
  llm: lc, // BaseChatModel
  maxModelTokens: 8192,
  thresholdPercent: 0.8, // summarize beyond 80% of context window
  minRecentMessages: 4,
});

See examples/LANGCHAIN_COMPRESS_HISTORY.md for a fuller copy-paste example.

API

Classes

  • TrimCompressor({ maxModelTokens?, thresholdPercent?, estimateTokens?, minRecentMessages? })
  • SummarizeCompressor({ model, maxModelTokens?, thresholdPercent?, estimateTokens?, minRecentMessages?, prompt? })

Interfaces

  • SlimContextMessage
  • SlimContextChatModel
  • SlimContextCompressor
  • SlimContextModelResponse

License

MIT

About

Lightweight, model-agnostic chat history compression (trim + summarize) for AI assistants.

Topics

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Contributors 3

  •  
  •  
  •