memory-mcp: Instructions file for Claude Code

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Build & Development Commands

bun install           # Install all dependencies (bun.lock is authoritative)
bun run build         # Remove dist/, then build TypeScript → JavaScript (tsc)
bun run test          # Run tests with coverage (vitest)
bun run typecheck     # Strict type checking (includes --noUnusedLocals --noUnusedParameters)
bun run typecheck:tests  # The same checks for src and tests/ (tsconfig.test.json; typecheck skips tests/)
bun run watch         # Watch mode for development
bun run clean         # Remove dist/ directory
bun run check:pack    # Fail when the npm tarball holds a build file that has no source
bun run check:bundle  # Fail when the committed bundle reports a version that differs from package.json
bun run check:bundle-content  # Fail when the committed bundle differs from a fresh build (run bun install first)

# Run a single test file
bunx vitest run tests/e2e/tools/entity-tools.test.ts

# Run tests matching a pattern
bunx vitest run -t "should create entities"

# Run server locally (after building) — Node is the shipped runtime
node dist/index.js

# Skip benchmark tests
# PowerShell:
$env:SKIP_BENCHMARKS=1; bun run test
# Bash/Unix:
SKIP_BENCHMARKS=1 bun run test

# Standalone tools (in tools/ directory — still npm + package-lock.json)
bun run tools:install # Install dependencies for all standalone tools
bun run tools:build   # Build all standalone tools

Architecture Overview

This is an MCP protocol wrapper around the @danielsimonjr/memoryjs library, exposing 241 knowledge graph tools via the Model Context Protocol. After the Phase 13 extraction, this repo contains only 5 TypeScript source files — all core graph logic lives in memoryjs (currently ^4.2.0). Phase 15 (v12.2.0) added 23 tools surfacing memoryjs v1.14+ features (bitemporal validity, OCC, RBAC, procedural memory, active retrieval, causal reasoning, world model). Phase 16 (v12.3.0) added 53 tools surfacing memoryjs v2.1.0 — do_not_remember exclusions (5), decision rationale + ADR markdown dual-write (10), structured project context (12), heuristic guidelines (10), tool affordance + ToolCallObserver pipeline (11), observation dedup (2), spell correction (3). Releases after Phase 16 added 12 more tools — a small "active project scope" pair (v12.3.2 backport) and a v12.5.0 engineering/diagnostics category. v12.7.0 upgraded to memoryjs 3.0.0 and added 16 tools — event memory (5), reconstructive memory (3 + 2 snapshot persistence), relation consolidation (2), agent reflection (4) — plus v3 graph-channel/explain options on hybrid_search, bringing the total to 241. Latest: v12.10.1 (plugin) — memoryjs ^4.2.0. loadGraph() now returns a read-only borrowed view (deep-frozen outside production), REST 4xx bodies use fixed messages, API keys are project-scoped, and RateLimiter gained TTL/LRU bounds. None of those surfaces are reachable from this wrapper except loadGraph(), whose contract is pinned by tests/unit/storage-readonly-contract.test.ts.