MemOS Roadmap

Public milestone plan. Each phase has a clear scope, success criteria, and a target timeline based on community adoption.

---

Phase 1 — Core Engine (v0.1) ✅ Released

Goal: A working, installable memory layer that any developer can use in 5 minutes.

Deliverables

• TypeScript SDK with MemOS class (store, retrieve, search, forget, summarize, link)
• Graph-based memory model (nodes + typed edges + metadata)
• SQLite storage backend with WAL mode and FTS5 full-text search
• Auto-linking via bag-of-words text similarity
• Extractive summarisation (fully local, no API calls)
• Python HTTP server (FastAPI) with full REST API
• CLI tool (memos command)
• Ollama adapter (Python)
• LangChain adapter (Python)
• Docker Compose deployment
• GitHub Actions CI (lint + test + typecheck)
• Comprehensive documentation (README, CONTRIBUTING, API reference)

Success criteria

• npm install @memos/sdk + 3 lines of code = working memory
• pip install memos + memos-server = HTTP server running
• All tests passing across Node 18/20/22 and Python 3.10-3.13
• Zero external dependencies for core functionality

---

Phase 2 — Semantic Search & Export (v0.2 → shipped in v1.6.26) ✅ Released

Goal: Make memory retrieval smarter and enable knowledge base integration.

Deliverables

• Embedding-based similarity search (local models via @xenova/transformers)
• Configurable embedding model (swap between local and API-based)
• Obsidian / Markdown export (memos export --format obsidian)
• Memory expiration (TTL) with automatic cleanup
• Memory tagging system (custom tags beyond type)
• Grafana-compatible metrics endpoint
• Performance benchmarks (10K, 100K, 1M memories)
• Backup / restore CLI commands

Success criteria

• Semantic search returns better results than FTS5 for conceptual queries
• Obsidian export produces linked markdown files with bidirectional links
• <10ms search latency at 100K memories on consumer hardware

---

Phase 3 — Multi-User & Plugin System (v0.3 → core adapters shipped in v1.6.26; backends/RBAC pending) 🚧 In progress

Goal: Enable production deployments with multiple users and custom backends.

Deliverables

• Multi-user isolation (namespace per user/agent)
• Role-based access control (read/write/admin)
• Plugin system for custom storage adapters
• PostgreSQL storage backend
• Redis storage backend (hot cache layer)
• Qdrant storage backend (vector search)
• Memory access audit log
• Rate limiting per user
• WebSocket API for real-time memory updates
• CrewAI adapter
• Vercel AI SDK adapter

Success criteria

• Multiple users can use the same MemOS server without data leakage
• Plugin authors can implement a storage backend in <100 lines
• PostgreSQL backend passes the same test suite as SQLite

---

Phase 4 — Production Hardening (v1.0 → partial: consolidation, semantic search, trust scoring shipped in v1.6.26; infra pending) 🚧 In progress

Goal: A battle-tested memory layer ready for production AI applications.

Deliverables

• Stable public API (no breaking changes until v2.0)
• Comprehensive test suite (90%+ coverage)
• Load testing and performance tuning
• Memory compression (deduplication, merging near-duplicates)
• Conflict resolution for concurrent writes
• Schema migration system
• Admin dashboard (web UI)
• Kubernetes Helm chart
• Cloudflare Workers adapter (edge deployment)
• Comprehensive architecture documentation
• Security audit

Success criteria

• 99.9% uptime in production deployments
• <5ms p99 latency for single-node reads
• Full backward compatibility for all 0.x APIs
• Published security best practices guide

---

Community milestones

Milestone	Target
100 GitHub stars	Week 1
500 stars	Month 1
2,000 stars	Month 2
5,000 stars	Month 3
First community adapter merged	Month 1
First blog post from a user	Month 2
First production deployment story	Month 3

---

How to influence the roadmap

1. Star the repo — signals demand, attracts contributors
2. Open an issue — feature requests, bug reports, use cases
3. Start a discussion — architecture proposals, integration ideas
4. Submit a PR — code speaks louder than issues
5. Share your usage — blog posts, tweets, conference talks

The roadmap is a living document. If your use case isn't covered, open an issue and let's talk.