arkiv

Universal personal data format. JSONL in, SQL out, MCP to LLMs.

The Format

Every record is a JSON object. All fields optional.

{"mimetype": "text/plain", "content": "I think the key insight is...", "uri": "https://chatgpt.com/c/abc", "timestamp": "2023-05-14T10:30:00Z", "metadata": {"role": "user", "conversation_id": "abc"}}
{"mimetype": "audio/wav", "uri": "file://media/podcast.wav", "timestamp": "2024-01-15", "metadata": {"transcript": "Welcome to...", "duration": 45.2}}
{"mimetype": "image/jpeg", "uri": "file://media/photo.jpg", "metadata": {"caption": "My talk at MIT"}}

The Stack

JSONL directory (human-readable, portable, durable)
       ⇅ arkiv convert
SQLite database (queryable, efficient, standard SQL)
       ↓ arkiv mcp
MCP server (tools → any LLM)

The two forms (directory and database) are isomorphic peers. arkiv convert goes either direction, auto-detected from input type.

Quick Start

pip install arkiv

# Point at a directory and query. arkiv.db is auto-created on demand.
arkiv query ./my-archive/ "SELECT content FROM records WHERE metadata->>'role' = 'user' LIMIT 5"

# Serve to any LLM via MCP
arkiv mcp ./my-archive/

# Explicit conversion (either direction)
arkiv convert conversations.jsonl archive.db              # JSONL → database
arkiv convert archive.db ./exported/                      # database → directory
arkiv convert archive.db 2024/ --since 2024-01-01         # temporal slice
arkiv convert archive.db archive.zip                      # pack for transport

Library: reading and writing bundles

Tools that produce or consume arkiv archives (the *-memex exporters and importers, for example) can use the bundle layer directly instead of reimplementing zip/tar handling:

from arkiv import write_bundle, records_to_jsonl, iter_jsonl, peek_jsonl

# Directory, .zip, or .tar.gz, chosen by extension. Byte-reproducible:
# identical records always produce identical bundle bytes.
write_bundle("export.zip", {
    "records.jsonl": records_to_jsonl(records),
    "schema.yaml": schema_bytes,
    "README.md": readme_bytes,
})

# Stream records from a directory, .zip, .tar.gz, or bare .jsonl / .jsonl.gz
# (members are read in place, never extracted). Non-object lines are skipped.
for record in iter_jsonl("export.zip", member="records.jsonl"):
    ...

first = peek_jsonl(path)  # None if absent or unreadable: for format detection

pack_bundle / unpack_bundle convert a whole directory tree (including media/) to and from a bundle; unpacking rejects path escapes and decompression bombs.

MCP Tools

Read-only by default. Start with arkiv mcp --writable db to enable the write tool.

Why

• Your data lives in silos (ChatGPT, email, bookmarks, photos, voice memos)
• Source toolkits (memex, mtk, btk, ptk, ebk) export it as JSONL
• arkiv gives you one format, one database, one query interface
• Any LLM can query it via MCP
• JSONL is human-readable and durable. SQLite is the most deployed database in history.

Spec and philosophy

• SPEC.md: full technical specification
• docs/PHILOSOPHY.md: why arkiv exists and how it composes with longecho