HN Showcase
Back to browse
GitHub Repository

Persistent memory for AI — powered by Shadow-Decay and HESTIA scoring

26 starsPython

Ember MCP – local persistent memory for LLMs, kills stale memories

by TimoLabs·Feb 20, 2026·1 point·0 comments
Visit ProjectView on HN

AI Analysis

●●●BangerBig BrainSolve My ProblemZero to One

Temporal memory with contradiction detection—Claude finally remembers job changes.

Strengths
  • •Shadow-decay engine automatically detects and flags stale information at write time
  • •100% local, zero cloud dependency with source-linked provenance
  • •Solves the genuine AI hallucination failure of contradictory facts accumulating
Weaknesses
  • •Still alpha with tuning overhead; token costs and weights need optimization per use case
  • •MCP ecosystem adoption is unproven; ecosystem lock-in risk
Target Audience

Claude Desktop users, developers building with MCP-compatible LLMs, knowledge workers

Similar To

Mem0 · Zappi · LlamaIndex persistent memory

Post Description

Most memory tools for LLMs just append to a flat list. Ember uses a different approach: a Shadow-Decay system that detects when new information contradicts stored memories and marks old ones stale automatically, rather than piling up conflicting facts.

It runs locally, stores everything in a vector database you control, and plugs into Claude (and other MCP-compatible clients) through 16 tools covering store, recall, graph search, and health monitoring.

The core problem it solves: you tell your AI you changed jobs, and three weeks later it still references the old one. Ember's contradiction engine catches that at write time.

Repo: https://github.com/Arkya-AI/ember-mcp

Still alpha — single user, some token overhead, scoring weights need tuning for different use cases. Posting here for technical feedback before wider release.

Similar Projects

Developer Tools●Mid

CogmemAi – Persistent Memory for Claude Code via MCP

Runs extraction and search server-side so your local MCP is a tiny HTTP client — no local DBs, no giant RAM leaks, and an easy npx install and .mcp.json or global MCP registration. It exposes clear tools (save_memory, recall_memories, extract_memories, get_project_context) and adds project-scoped + global preferences — a pragmatic fix for Claude Code's tiny flat-file memory. The tradeoff is obvious: usefulness depends on the hosted API (privacy, uptime, cost), and the repo looks early-stage with minimal commits and docs beyond the quickstart.

Niche GemShip ItSolve My Problem
hifriendbot
203mo ago