Kopeng banner
djy89 djy89

Kopeng

Development community

Description

Persistent, self-curating memory for coding agents. MCP tools + REST API, 100% local inference, with an autonomous nightly consolidation pass that collapses duplicates, decays stale memories, and routes contradictions - every mutation snapshot-first, audited, and reversible.

Installation

This entry records only its repository, not the path inside it, so there is no exact command to give. Open the source below and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

README

KOPENG

[![CI](https://github.com/djy89/kopeng/actions/workflows/ci.yml/badge.svg)](https://github.com/djy89/kopeng/actions/workflows/ci.yml) ![License: BUSL-1.1](https://img.shields.io/badge/license-BUSL--1.1-blue) ![Node >= 20](https://img.shields.io/badge/node-%3E%3D20-brightgreen) ![TypeScript strict](https://img.shields.io/badge/TypeScript-strict-3178c6) ![Local-first](https://img.shields.io/badge/inference-100%25%20local-orange)

**Persistent, self-curating memory for coding agents — it doesn't just remember what you've done, it cleans up after itself, fully local.**

KOPENG 0.x is a **local developer preview**: a self-hosted memory system for a single expert developer, on one machine, bound to loopback. Every autonomous layer ships OFF and is labeled advanced. What backs reliability is engineering rigor, not scale — a zero-LLM pinned-clock replay regression net, adversarial GATE reviews run against a *copy* of real data, idempotent locked consolidation passes, and fail-open/fail-silent behavior everywhere a hook or service could stall.

Renamed 2026-07 to its current codename. Everything now uses `kopeng` — hook env vars are `KOPENG_*` and the client data directory is `~/.kopeng/`.

KOPENG is a memory and context layer for coding agents (Claude Code, Codex CLI), exposed as MCP tools plus a REST API. It learns from observed tool-use — passively turning repeated tool calls, error-then-fix patterns, hot files, and cross-session sequences into confidence-scored memories with no LLM cost — and serves them back through a hybrid retrieval pipeline (RRF fusion of semantic + keyword, optional cross-encoder rerank, confidence-blended ranking), all running on local quantized ONNX models so there is no per-query API cost or data egress. Its distinguishing layer is operator-gated nightly consolidation (the "dreaming Librarian"): a deterministic-first engine that collapses duplicate memories, decays stale ones, and routes contradictions and supersessions — every mutati