feat(providers): add Basic Memory provider

[groksrc]

Summary

Adds a Basic Memory provider to memorybench (GitHub) — an open-source, local-first knowledge graph that stores information as Markdown files and exposes full-text + semantic (vector) search via a CLI and MCP server.

This PR implements the MemoryProvider interface for Basic Memory, closing basicmachines-co/basic-memory#672.

What changed

• src/providers/basic-memory.ts — new provider that drives the local bm CLI via node:child_process.spawn (matching the pattern used by the filesystem and rag providers).
• Isolation: each benchmark containerTag maps to a throwaway BM project rooted under data/providers/basic-memory/ (gitignored). The provider redirects BASIC_MEMORY_CONFIG_DIR / BASIC_MEMORY_HOME so it never touches the user's real config or projects.
• Ingest: sessions are written as Markdown notes with speaker-attributed dialogue via bm tool write-note.
• Indexing: polls bm status until sync settles, then runs bm reindex --full --embeddings to ensure vector embeddings are up to date (lazy indexing otherwise causes semantic search misses).
• Search: uses bm tool search-notes --hybrid (FTS + semantic). BM's default FTS uses implicit-AND which fails on natural-language questions; hybrid mode correctly handles them.
• Registration: provider wired into providers/index.ts, types/provider.ts (ProviderName), utils/config.ts (no API key required).
• Docs: provider table in src/providers/README.md updated with a Basic Memory Setup section.

Setup (for reviewers)

# Install Basic Memory CLI
pip install basic-memory   # or: brew install basic-memory

# Run benchmark
bun run bench --provider basic-memory

The CLI path is auto-resolved at init; override with BASIC_MEMORY_CLI env var or config.baseUrl.

---

🤖 Generated with Claude Code

[phernandez]

Checked this against current Basic Memory main (368e607622e9af3d982d0429cb48cfc6c83521f1) in a temporary config/project. Project creation and the exact stdin tool write-note command work and return JSON as expected.

One compatibility fix is needed in awaitIndexing: settled bm status now prints indexed and settled, not no changes. The current regex therefore waits out the full 120-second loop even when indexing is already idle. Prefer bm status --project PROJECT --local --json --wait --timeout 120 and check its exit status, or parse readiness.phase (idle) from JSON. The structured result also exposes files, relations and embeddings stages.

Please also propagate a failed embeddings rebuild rather than logging it and reporting every document completed. Similarly, returning [] after a failed search turns a provider/configuration failure into a retrieval miss in the benchmark score. Those should fail the run or be explicitly recorded as provider failures.

I have verified CLI creation/write/status compatibility only; I have not run a scored benchmark or validated hybrid quality. Keeping basicmachines-co/basic-memory#672 open until the provider lands and an end-to-end smoke run verifies ingest, indexing, search and cleanup.