Appendix E中文

mini-mem Source Listing

The complete project is in examples/ch13-mini-mem/ (v1.1.0). Below are the paths and responsibilities of its key files. claude plugin validate plugin passes with zero warnings.

Project Layout


examples/ch13-mini-mem/
├── package.json                # dependencies and npm scripts (Node.js 20.12+)
├── tsconfig.json
├── README.md
├── src/
│   ├── hooks/
│   │   ├── save-hook.ts        # PostToolUse / PostToolUseFailure: capture observations
│   │   ├── context-hook.ts     # SessionStart: inject the index, skip on resume
│   │   ├── context-render.ts   # index rendering + trimming to the 10,000-character budget
│   │   └── budget-test.ts      # builds oversized data to verify budget trimming
│   ├── mcp/
│   │   ├── server.ts           # MCP server: search + get_observations (read-only annotations)
│   │   └── test-client.ts      # tests over stdio using the SDK Client
│   ├── db/
│   │   ├── schema.ts           # tables + FTS5 + triggers (the update trigger watches only indexed columns)
│   │   └── store.ts            # read/write wrapper: busy_timeout, safe MATCH, LIKE fallback for CJK
│   └── utils/
│       ├── stdin.ts            # reads JSON from stdin
│       ├── project.ts          # project identification
│       └── title-extractor.ts  # rule-based titles and types; skips meaningless tools
└── plugin/
    ├── .claude-plugin/
    │   └── plugin.json         # plugin metadata (including author)
    ├── hooks/
    │   └── hooks.json          # SessionStart (sync) + PostToolUse / PostToolUseFailure (async)
    └── .mcp.json               # MCP server registration

Running It


cd examples/ch13-mini-mem
npm install
npm run build
npm run test:save      # simulate one successful tool call → written to SQLite
npm run test:fail      # simulate one failed tool call
npm run test:context   # simulate SessionStart(startup) → prints the index
npm run test:resume    # simulate SessionStart(resume) → nothing injected
npm run test:budget    # verify the 10,000-character budget
npm run test:search    # start the MCP server, list the tools, and call them
npm run claude:dev     # claude --plugin-dir ./plugin, load it in Claude Code

Data goes to ~/.mini-mem/ by default. For tests, point MINI_MEM_DATA_DIR at a temporary directory.

Mapping to claude-mem

mini-mem file	claude-mem v13.35.0 counterpart	What’s simplified
`save-hook.ts`	`src/cli/handlers/observation.ts`	No spool or worker; the hook writes SQLite directly (and relies on `async` to avoid blocking the host)
`context-hook.ts`	`src/cli/handlers/context.ts`	No pre-rendered cache; queries fresh every time
`context-render.ts`	`src/services/context/ContextBudget.ts`	Trims by count only, with a simpler trimming order
`server.ts`	`src/servers/mcp-server.ts`	2 tools, without `mem_search`’s stepwise protocol
`store.ts`	`src/services/sqlite/`	A single file, no migrations
`title-extractor.ts`	Generated by the observer model	Rule-based extraction, no model call

Where to Go Next

After finishing mini-mem, extend it in this order (Chapter 14 goes into detail):

Go asynchronous: hooks only write to disk and hand off to a background process (Chapter 6’s spool pattern)
Generate observations with a model: replace rule-based extraction with Claude (ai-compressor.ts in examples/ch14-extensions)
Add a timeline tool: fetch context before and after a point in time, taking care with ordering when several records share a timestamp
Add vector search: Chroma; for non-English memory, switch to a multilingual embedding model
Add a Viewer: Express + SSE