#2121st of 30 in Memory

reme

File-based Markdown memory store for AI agents with BM25 search

Stars
3.8k
License
Apache-2.0
Last commit
Nov 2026
Last release
Nov 2026
Language
Python
Stars, last 29 weeks+1,029
25 Apr 2026: 2,785 stars2 May 2026: 2,841 stars9 May 2026: 2,882 stars16 May 2026: 2,919 stars23 May 2026: 2,958 stars30 May 2026: 3,001 stars6 Jun 2026: 3,028 stars13 Jun 2026: 3,058 stars20 Jun 2026: 3,080 stars27 Jun 2026: 3,105 stars4 Jul 2026: 3,129 stars11 Jul 2026: 3,160 stars18 Jul 2026: 3,184 stars25 Jul 2026: 3,204 stars1 Aug 2026: 3,238 stars8 Aug 2026: 3,273 stars15 Aug 2026: 3,306 stars22 Aug 2026: 3,327 stars29 Aug 2026: 3,359 stars5 Sept 2026: 3,409 stars12 Sept 2026: 3,443 stars19 Sept 2026: 3,483 stars26 Sept 2026: 3,522 stars3 Oct 2026: 3,548 stars10 Oct 2026: 3,573 stars17 Oct 2026: 3,642 stars24 Oct 2026: 3,689 stars31 Oct 2026: 3,736 stars7 Nov 2026: 3,775 stars14 Nov 2026: 3,814 stars

2,785 on 25 Apr 20263,814 on 14 Nov 2026

From GitHub's daily star counts.

Checked on GitHub.

Overview

ReMe stores agent memory as plain Markdown files with frontmatter and wikilinks, organized into session, daily and digest layers. It runs as a local service (CLI, HTTP, MCP or embedded Python API) that indexes files with BM25 and optional embeddings, and uses an LLM to consolidate conversations and resources into long-term notes. Plugins exist for Claude Code, Codex, Qoder, OpenClaw, Hermes and DeepSeek Harness.

Who it is for: Developers giving coding or personal agents shared, file-based long-term memory

Strengths

  • Memory is ordinary Markdown; indexes under metadata/ are rebuildable from the files
  • File operations, BM25 search and wikilink traversal work without LLM credentials
  • Integrations for Claude Code, Codex, Qoder, OpenClaw, Hermes, plus CLI, HTTP, MCP and Python
  • Embeddings are off by default, so no embedding model or key is needed to start

Weaknesses

  • auto_memory, auto_resource, auto_dream and proactive refresh require an external LLM API key
  • Semantic search requires editing default.yaml to enable embedding components
  • Requires running a separate long-lived service (reme start) alongside the agent
  • Source install needs Node.js 22.13+ to build the Studio static assets

What it needs

  • no GPU
  • Docker + Compose
  • Needs Python 3.11+, OpenAI-compatible LLM API (for auto features), Optional embedding API, Docker Compose 2.24.0+ (Docker path)
  • Models: OpenAI-compatible LLM endpoints (e.g. DashScope), Optional OpenAI-compatible embeddings
  • port 2333

What changed

  1. “refactor(agent)!: unify lossless streaming and collected replies by @jinliyl in https://github.com/agentscope-ai/ReMe/pull/713”

  2. “refactor(codex)!: share app-server processes across agent wrappers by @jinliyl in https://github.com/agentscope-ai/ReMe/pull/709”

From the release notes. No flag means the note does not say so, not that the upgrade is safe.

Also in Memory

See all 30
Also in Memory
RankProjectScore
1hindsightAgent memory server with retain, recall and reflect operations61.7k starsMIT95 out of 100
2gbrainMemory store with sources and a knowledge graph for coding agents33.2k starsMIT93 out of 100
3MemPalaceLocal verbatim memory store for AI agents with semantic search and MCP60.6k starsMIT90 out of 100
#4SupermemoryMemory and context engine for AI agents, with user profiles and hybrid search31.7k starsMIT89 out of 100
#5Mem0Long-term memory layer that stores and retrieves user facts for LLM agents Live demo ↗ (opens in a new tab)69.2k starsApache-2.088 out of 100