letta-ai/letta-filesystem-to-memfs

Letta Filesystem to MemFS

Use this skill when a user wants the old Letta Filesystem behavior: upload documents, chunk them, attach them to an agent, and let the agent search/open them.

MemFS is not the same product. It is git-backed markdown memory. To mimic the old workflow, use the bundled CLI:

1. Extract PDFs/docs to markdown text.
2. Chunk the text into stable markdown files under documents/<corpus>/<doc>/chunks/.
3. Write a small pinned index under system/filesystem/<corpus>.md.
4. Index only the corpus chunk files in QMD for semantic search.
5. Review the MemFS git diff. Commit only if asked.

Quick workflow

# Set this to the skill directory shown in the skill load header.
SKILL_DIR="/path/to/letta-filesystem-to-memfs"

# From any directory. MEMORY_DIR should point at the target agent's memfs repo.
uv run --with pymupdf \
  "$SKILL_DIR/scripts/letta_fs_to_memfs.py" ingest \
  --memory-dir "$MEMORY_DIR" \
  --corpus product-docs \
  --source ./docs/ \
  --source ./guide.pdf \
  --source https://arxiv.org/pdf/2310.08560

cd "$MEMORY_DIR"
git status --short
git diff --stat

Search the chunk corpus lexically:

uv run "$SKILL_DIR/scripts/letta_fs_to_memfs.py" search \
  --memory-dir "$MEMORY_DIR" \
  --corpus product-docs \
  "memory hierarchy" \
  -n 5

Semantic search with QMD. The CLI creates a corpus-scoped QMD collection over chunk files only:

uv run "$SKILL_DIR/scripts/letta_fs_to_memfs.py" qmd setup \
  --memory-dir "$MEMORY_DIR" \
  --corpus product-docs

uv run "$SKILL_DIR/scripts/letta_fs_to_memfs.py" qmd query \
  --memory-dir "$MEMORY_DIR" \
  --corpus product-docs \
  "memory hierarchy" \
  -n 5

Use qmd reindex after changing corpus files, and qmd search / qmd vsearch when you specifically want keyword-only or vector-only search.

Layout

The ingest script writes:

system/filesystem/<corpus>.md
  Pinned index and operating instructions for the corpus.

documents/<corpus>/manifest.md
  Corpus manifest.

documents/<corpus>/<doc-slug>/manifest.md
  Per-document metadata and chunk table.

documents/<corpus>/<doc-slug>/chunks/chunk-0001.md
  Chunk content with frontmatter metadata.

documents/<corpus>/chunks.jsonl
  Machine-readable chunk export for custom indexing or debugging.

Old API mapping

| Old Filesystem | MemFS mimic | |---|---| | folders.create | --corpus <name> creates documents/<corpus>/ | | folders.files.upload | letta_fs_to_memfs.py ingest --source <file-or-directory-or-url> | | OCR/chunk/embed job | Extract + chunk locally; qmd setup / qmd reindex for semantic search | | agents.folders.attach | Enable MemFS, then review and sync repo changes when appropriate | | open_file | Read chunk markdown files by path | | grep_file | rg or letta_fs_to_memfs.py search | | search_file | letta_fs_to_memfs.py qmd query over the corpus chunk collection |

Heuristics

• Use system/filesystem/<corpus>.md for the small always-visible index only.
• Do not pin full documents into system/; it will bloat the prompt.
• Keep chunks outside system/, usually under documents/<corpus>/....
• Passing a directory to --source recursively ingests supported files (.pdf, .md, .txt, .json, .csv, .html, .xml).
• Use --glob / --exclude for messy directories. Defaults exclude .git, node_modules, .venv, and __pycache__.
• URL downloads default to --max-download-mb 100; set 0 for unlimited.
• Re-ingesting the same document slug replaces that document's old chunk directory, so stale chunks do not survive chunk-size changes.
• Use chunk sizes around 2,000-4,000 characters with 200-500 character overlap.
• Use the CLI's qmd subcommands when the user needs semantic search over many chunks.
• Preserve source URLs, checksums, page markers, chunk numbers, and corpus names in the generated files.

CLI reference

uv run --with pymupdf "$SKILL_DIR/scripts/letta_fs_to_memfs.py" ingest --help
uv run "$SKILL_DIR/scripts/letta_fs_to_memfs.py" search --help
uv run "$SKILL_DIR/scripts/letta_fs_to_memfs.py" qmd setup --help
uv run "$SKILL_DIR/scripts/letta_fs_to_memfs.py" qmd query --help

Compatibility wrappers remain for older examples:

uv run --with pymupdf "$SKILL_DIR/scripts/ingest_documents.py" --memory-dir "$MEMORY_DIR" --corpus docs --source ./docs
uv run "$SKILL_DIR/scripts/search_corpus.py" --memory-dir "$MEMORY_DIR" --corpus docs --query "refund policy"

PDF notes

The ingest script uses PyMuPDF when it sees a PDF. Run it with uv run --with pymupdf ....

For scanned PDFs or complex tables, use the tools/extracting-pdf-text skill first, then ingest the extracted markdown/text file with this skill.

See references/design.md for design notes and edge cases.