Summary

Flip the primary install command across all first-touch docs to uv tool install 'memtomem[all]'. Keep the bare uv tool install memtomem shape as a secondary "Minimal install" option. Update the cache-lag blockquotes (from #393) to reference the new primary command.

5 files, +33 / −14 lines. Docs-only, no code changes.

Closes #391.

Why

The [all] extra (memtomem[onnx,ollama,openai,korean,code,web]) bundles exactly the features the public docs advertise:

• ONNX dense embeddings — required by both "Recommended" presets (English / Korean-optimized) that mm init pushes by default
• Korean tokenizer (kiwipiepy) — advertised as a multilingual selling point
• Ollama / OpenAI providers — listed in every embeddings table
• Code chunker (tree-sitter) — implied by "index markdown + code"
• Web UI (fastapi) — surfaced in its own README section + mm web command

A fresh uv tool install memtomem ships none of these. Observed on 2026-04-23 in a clean tool env:

fastembed                 MISSING ✗
onnxruntime               MISSING ✗
kiwipiepy                 MISSING ✗
tree_sitter               MISSING ✗
fastapi                   MISSING ✗
ollama                    MISSING ✗
openai                    MISSING ✗

Runtime behaviour after picking a feature that needs a missing extra:

• mm web → loud error with remediation (uv tool install --reinstall 'memtomem[web]')
• mm init -y --provider onnx → wizard accepts, writes provider=onnx (0d); index/search then warn "degraded mode — Run 'mm embedding-reset' to fix"
• mm init -y --tokenizer kiwipiepy → wizard accepts; search-time warning "kiwipiepy not installed — falling back to unicode61"

Users aren't silent-degraded at runtime (that framing from the original issue was revised in this fact-check comment), but the install command doesn't match what the same README promises. Users who follow the README literally can end up configuring a preset that never reaches dense search and only learn about it via runtime warnings.

Going with Option 1 from the issue discussion: the recommended setups already require [all]-level extras, so making the primary command deliver that is consistent with the docs' implicit promise. Users who specifically want a ~40 MB lightweight install can take the Minimal option.

Files changed

Intentional residual uv tool install memtomem bare lines (4 total) are all inside the Minimal install alternatives, clearly labeled as such.

Scope — intentionally out

• Wizard-time validation of extras (e.g. mm init -y --provider onnx refusing when fastembed is missing) — separate gap, tracked adjacent to mm init: auto-install consistency for python extras + mismatch banner (Phase 2 of #360) #362.
• [all] packaging composition — no pyproject.toml changes in this PR. If we later split into a lighter "recommended" extra set, the docs can re-point.
• Absolute URLs vs relative links — the PyPI README already mixes both (see line 33); kept the same style.

Test plan

• Verified on a fresh uv tool install 'memtomem[all]' in a clean env: mm --version prints 0.1.24, fastembed, kiwipiepy, fastapi, ollama, openai, tree_sitter all import.
• Inline alternative uv tool install memtomem verified in a prior run (48 MB, BM25-only default config).
• Anchor #minimal-install added via explicit <a id> tag (safer than heading-derived slug).
• Cross-reference #option-c-from-source-for-development-or-testing matches GitHub's default anchor slug for the existing ### Option C: From source (for development or testing) heading.
• No lint/test failures expected — docs only.
• Reviewer: render each page and confirm the Minimal/alternative blocks read well; verify [all] shows up unquoted-unbracket-exploded in rendered code blocks.

🤖 Generated with Claude Code