claudiodrews-memory-os

name	claudiodrews--memory-os
description	Hệ thống bộ nhớ 7 lớp chạy local dành cho AI agent — SQLite + Qdrant + Fabric, zero cloud dependency. Agent nhớ project, decision, reasoning pattern xuyên session.
allowed-tools	Bash, Read, Write
user-invocable	true

Memory OS biến Hermes Agent (hoặc bất kỳ LLM agent nào) thành collaborator dài hạn: nhớ đủ thứ giữa các session mà không cần cloud, không vendor lock-in.

7 Lớp Kiến Trúc

Layer 1 — Workspace Files
  MEMORY.md, USER.md, CREATIVE.md inject vào system prompt mỗi turn

Layer 2 — Session History
  SQLite + FTS5 full-text search — archive toàn bộ conversation

Layer 3 — Structured Facts
  Entity resolution + trust scoring + auto feedback loop

Layer 4 — Fabric (Cross-Session Extraction)
  LLM-powered summarization, 16 tools: fabric_recall, fabric_write, fabric_brief

Layer 5 — Vector Database
  Qdrant 4096-dim cosine + BM25 sparse, fallback: hybrid→dense→lexical→SQLite
  Weekly decay scan + semantic dedup (cosine > 0.92)

Layer 6 — Auto-Curated Wiki
  Self-organizing knowledge vault: concepts, entities, comparisons → Qdrant

Layer 7 — Ground Truth Hierarchy
  SOUL.md + rulebook.md — injected memory = authoritative
  Ngăn agent re-discover context đã có → tiết kiệm token

Install

# One-command setup — Docker orchestrate Qdrant + Redis + ARQ worker
curl -fsSL https://raw.githubusercontent.com/ClaudioDrews/memory-os/main/install.sh -o /tmp/memory-os-install.sh
# Inspect first: head -40 /tmp/memory-os-install.sh — then run if safe:
bash /tmp/memory-os-install.sh

Core Tools (Fabric Layer)

fabric_recall(query, n=5)     # semantic search memory
fabric_write(content, tags)   # lưu fact mới
fabric_brief(session_id)      # summarize session → store

Tích Hợp

# Compatible với bất kỳ LLM provider nào Hermes support
# OpenRouter, OpenAI, Anthropic, Ollama, local models

# Databases:
# state.db — session history
# memory_store.db — facts + HRR (hierarchical relation)
# Qdrant — semantic vectors

So Với Alternatives

Feature	Memory OS	mem0	Zep	Letta
Fully local	✅	❌	❌	❌
Ground Truth layer	✅	❌	❌	❌
Trust-scored facts	✅	❌	✅	❌
Self-curating wiki	✅	❌	❌	❌
No subscription	✅	❌	❌	❌

Layer 7 — Quan Trọng Nhất

Nếu không có Ground Truth ranking: agent bỏ qua injected context, gọi API tìm lại thứ đã có trong prompt → waste token + slow. Layer 7 force agent treat injected memory là authoritative source.

Source

https://github.com/ClaudioDrews/memory-os · MIT