Collection basis: the GitHub Trending bot issue happydog-intj/ai-xiaohongshu-daily #578 (“GitHub Trending AI 早报 · 2026-10-20”, Top 8), compared with #575 (Oct 19), ranked by stars gained today. Seven of the eight repos already have more than 40,000 stars (MadsLorentzen/ai-job-search 56.5k, anomalyco/opencode 217.6k, ComfyUI 139.5k, ruvnet/ruflo 77.0k, stablyai/orca 98.0k, headroomlabs-ai/headroom 76.1k, HKUDS/CLI-Anything 55.1k) and were excluded under our rules.
akitaonrails/ai-memory — +666 ⭐ today (2 days in a row)
URL: https://github.com/akitaonrails/ai-memory
Total stars: 11,162 · Change vs. yesterday: +513 → +666 (about +30%) · Language: Rust · License: MIT
ai-memory is a long-term memory layer for AI coding agents. It records what happens in agent sessions, turns it into wiki pages, and hands the next session, in any agent and on any machine, a brief plus a searchable history. A single Rust binary runs an MCP/HTTP server that owns one data directory; lifecycle hooks post observations to it, and a single SQLite writer compiles them into markdown pages (capture → consolidate → recall → handoff). The source of truth is an ordinary git-versioned .md wiki, while SQLite (FTS5, entities, embeddings) is just a rebuildable index, so there is no vector store to babysit. Retrieval fuses full-text search, entity matching and graph-neighbor expansion with RRF, plus optional vector search. It integrates with 18+ agents including Claude Code, Codex, Cursor, Gemini CLI, OpenCode and Kimi Code, and it is riding the rise of developers who mix several agents and need memory that crosses vendors.
Key features:
- Cross-agent handoffs: “a protocol here, not a convention” — typed, owned, and claimed exactly once
- Cross-machine continuity: memory lives on a server you run, so work resumes on another machine
- Team mode: per-project shared knowledge, personal handoffs, multi-user auth, attribution and an audit log
- Silent capture with a privacy boundary: sanitized before storage, with per-repo
[capture]exclude or allowlist rules - Zero-LLM default: capture, search and handoffs need no API key; an LLM only improves summaries and semantic search
- Two-command setup per agent (
install-mcp,install-hooks) with backups and clean uninstall; AUR, Docker and macOS builds
Practical takeaway: teams that run several coding agents can keep project context and work handoffs on one internal server, cutting the cost of re-explaining context when people or tools change.
Tags: #AgentMemory #MCP #Rust #ClaudeCode #CrossAgentHandoff #CodingAgents
Leave a Reply