Appendix F中文
References
Core Projects
claude-mem
- GitHub: github.com/thedotmack/claude-mem. This book’s baseline is v13.35.0 (2026-10-08), Apache-2.0 (since v13.0.0; AGPL-3.0 before that)
- Website and docs: claude-mem.ai
- Interactive intro course: cmem.ai/learn
- Install:
npx claude-mem install - Design docs in the source worth reading:
docs/ip-boundary.md(the open-source boundary),docs/server.md(the Server runtime),docs/native-memory-bridge.md(the native memory bridge),docs/native-prompt-identity.md(turn identity),workers/sync-hub/DEPLOY.md(deploying the sync hub). Note that some docs, such asdocs/architecture-overview.md, lag behind the code; trust the code
Claude Code
- Hooks reference: code.claude.com/docs/en/hooks
- Plugins:
claude plugin validate,claude --plugin-dir(verified for this book on 2.1.296) - Memory: code.claude.com/docs/en/memory
Model Context Protocol
- Specification: modelcontextprotocol.io
- TypeScript SDK: github.com/modelcontextprotocol/typescript-sdk (this book’s examples use 1.32)
Papers and Research
Lost in the Middle
Liu, N. F., et al. (2024). Lost in the Middle: How Language Models Use Long Contexts. Transactions of the Association for Computational Linguistics, 12, 157–173.
Models are more likely to miss information in the middle of a long context. Chapter 3 draws on this result when discussing context rot.
ACT-R and the Power Law of Forgetting
Anderson, J. R., & Schooler, L. J. (1991). Reflections of the Environment in Memory. Psychological Science, 2(6), 396–408.
How retrievable a memory is decays with time according to a power law and strengthens with each repetition. claude-mem’s reinforcement ranking (src/services/reinforcement/) is computed from this model with a decay exponent of 0.5. Chapters 7, 10, and 20.
The Ebbinghaus Forgetting Curve
Ebbinghaus, H. (1885). Über das Gedächtnis: Untersuchungen zur experimentellen Psychologie. Leipzig: Duncker & Humblot.
Chapter 20 compares an exponential-decay retention score with ACT-R’s power-law decay.
BM25 Ranking
Robertson, S., & Zaragoza, H. (2009). The Probabilistic Relevance Framework: BM25 and Beyond. Foundations and Trends in Information Retrieval, 3(4), 333–389.
SQLite FTS5’s rank is based on BM25. Chapter 7 and Appendix D.
Retrieval-Augmented Generation (RAG)
Lewis, P., et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS 2020.
Chapters 1 and 8 compare progressive disclosure with traditional RAG.
Open-Source Projects
Comparable Agent Memory Systems (Chapter 15)
| Project | Summary | Where |
|---|---|---|
| Mem0 | Multi-level memory architecture, with user / session / agent levels | mem0.ai |
| Zep / Graphiti | Temporal knowledge graph | getzep.com |
| Letta (formerly MemGPT) | OS-style virtual context management; the agent reads and writes its own memory | letta.com |
| Cognee | Knowledge-graph-enhanced memory | cognee.ai |
| LangGraph | Short-term state + a long-term Store | langchain-ai.github.io/langgraph |
Infrastructure
| Project | How this book uses it |
|---|---|
| bun:sqlite / better-sqlite3 | claude-mem uses the former; this book’s examples use the latter (12.x, which supports Node 24) |
| SQLite FTS5 | Full-text search, sqlite.org/fts5.html |
| Chroma | Vector search; claude-mem’s default local embedding is all-MiniLM-L6-v2 |
| PostgreSQL / Valkey / BullMQ | Storage and queues for claude-mem’s Server runtime (Chapter 16) |
| pg_trgm | Speeds up CJK substring search on Postgres (tested in Chapter 16) |
| Cloudflare Durable Objects | The per-user ordered log in claude-mem’s sync hub (Chapter 19) |
Further Reading
Context Engineering and Agent Design
- Building Effective Agents (Anthropic Engineering, 2024): agent design patterns
- Effective Context Engineering for AI Agents (Anthropic Engineering, 2025): context rot, the “right altitude” for a system prompt, just-in-time context, and compaction / structured notes / subagents for long tasks, all discussed in Chapter 3, have counterparts in this article
Databases
- SQLite Is Not a Toy Database (Anton Zhiyanov): what SQLite can and can’t do in production
- The transactional outbox pattern: the write path of Chapter 16’s Server runtime and Chapter 19’s “the database is the queue” are both variants of it
From the Same Series (in Chinese)
- The source-reading books on DeepSeek Harness and OpenClaw cover two of the hosts in Chapter 18
- AI Agent Evals in Practice: how to measure the effect of memory with controlled experiments (Chapter 17)
Community and Feedback
This book is open source on GitHub and readable at inferloop.dev, under the CC BY-NC-SA 4.0 license:
- Errata and discussion: open an issue or PR on GitHub
- Reposting: credit the source; no commercial use
- Commercial licensing: contact inferloop.dev
The code examples (the
examples/directory) are MIT-licensed and free to use in commercial projects.