context-compression

Rolling context compression for Claude Code: a zero-dependency proxy that summarizes old messages and keeps recent turns verbatim — beat the context-window limit and cut token & cache costs. Turn-aware keep policy, concurrency-hardened.

Deterministic, auditable LLM context compression — regex whitelist guarantees structured facts (IDs, URLs, dates, code) survive. Two layers: retention + heuristic knapsack.

Unified compression pipeline for LLM inputs: trim irrelevant rows and fields, re-encode the rest in the cheapest lossless format (TOON / JSON / CSV, measured on your tokenizer), and account for every token saved with an audit trail of what was removed. 95% fewer tokens on realistic payloads, 100% needle recall.

A research-oriented framework for adaptive context engineering in RAG systems, exploring structured representation, compression, and progressive context expansion through controlled experiments.RAG

Lightweight Terminal Intelligent Context Compression Engine — Smart compression for tool output, logs, and RAG chunks, reducing 60-95% token consumption

Semantic context triage for AI agents — classify before compression so constraints, exact values, and decision reasoning survive long sessions

Official implementation of AsymSpec: Context-Asymmetric Speculative Decoding for Agentic LLMs (EMNLP 2026)

Compress chat histories into minimal JSON graphs to reduce token costs and maintain context across different LLM sessions.

Up to 99% token savings on AI agent context. Free. Provider-agnostic (Claude, Cursor, Codex, Cline, any MCP agent). 91.62% over 626.8M auditable tokens. Try: nuxs.ai/playground

A context engine for DeepSeek Harness — reversible, token-budgeted compression of the live context window.

Local-first AI computer control plane: governed model routing, durable memory, context compression, bounded agents, MCP, and receipt-backed execution on Windows and network AI boxes.

TokenPack packs long documents, codebases, PDFs, and folders into compact, evidence-dense LLM context using local embeddings, evidence scoring, and budget-aware selection.

Compress long contexts into a sparse latent graph that a frozen LLM reads as a soft prompt. Research code with tests, controls and honest reporting.

Quality-aware context optimization for AI agents: compress with PariTok, preserve exact facts, validate outcomes, and fall back safely.

Local-first context compression for AI agents — content-aware, reversible, zero runtime deps. Cuts agent token usage 40-95% while keeping the signal.

The neural insulation layer for AI coding agents. 93% token reduction, 6× longer sessions. Serena + Semble + Headroom + RTK + mem0 in one installer.

LLM token compression via images: a transparent Antrhopic/OpenAI proxy that renders agent context (system prompt, tool docs, tool output, history) to pixels, so coding-agent CLIs like OpenCode send far fewer input tokens. Keeps tool calls and multi-turn intact. ~39% fewer tokens on real runs.

⚡ Absolute command over your AI agent's token budget — local-first compression of noisy tool output before it reaches the model. Claude Code / OpenCode / Codex / Gemini.

Local-first adaptive context compiler and privacy gateway for LLMs and agents