daidocs (Python)

A pure-Python reader for DaiDocs .dai memory stores, plus a small daidocs command that forwards to the Node engine.

DaiDocs is an open plain-text format for AI memory. A store is a folder of .dai files plus an _index/. This package gives Python code a way to read that store. It does not reimplement the engine that writes stores: writing, extraction, embeddings and ranking stay in the Node project.

Requirements

• Python 3.8+
• Node 18 or newer (required). DaiDocs is a Node project; this package drives that engine. If Node is missing when you first run daidocs, the command tells you it is required and offers to install it for you (via your OS package manager, such as winget on Windows, after you confirm). You can also install Node (LTS) yourself from https://nodejs.org/ . The full setup needs a system Node, because Claude Code and Claude Desktop launch it themselves.

Verify Node with:

node --version

Install

pip install daidocs

This pulls in pyyaml. Node is a separate prerequisite (see Requirements): the daidocs command drives the published npm engine via npx, so Node must be present for setup, convert, capture and the MCP server.

Reader (pure Python)

from daidocs import Store

store = Store("~/DaiDocs")        # a store folder, or a project folder

store.manifest()                  # list of manifest entries (dicts)
store.read("chat_20260807_d1dd87da")
                                  # dict: id, path, header, understanding,
                                  #       segments, raw
store.facts()                     # parsed _index/facts.jsonl rows
store.events()                    # parsed _index/events.jsonl rows
store.profile()                   # parsed _index/profile.jsonl rows
store.search("deploy")            # substring filter over the manifest

read() returns the document's three parts:

• header: the YAML frontmatter (id, type, title, lang, summary, tags, ...).
• understanding: the JSON extraction block under # Understanding.
• segments: the ## [seg n/N] blocks under # Content, each as {"n", "total", "text"}.
• raw: the full original file text, preserved losslessly.

You can also parse a .dai string directly without a store:

from daidocs import parse_dai
parsed = parse_dai(open("chat_20260807_d1dd87da.dai", encoding="utf-8").read())

CLI (drives the Node engine, Node required)

daidocs setup
daidocs convert

These run npx -y daidocs@latest <args> underneath, so they behave exactly like the Node engine. Node 18+ is required: if Node is missing or older, the command says so and offers to install it for you (after you confirm), rather than silently falling back. Once Node is present, run daidocs setup to get the full setup: Claude Code hooks, the Claude Desktop MCP server, and conversion.

Note: if you also have the npm package installed globally, both provide a daidocs command and can shadow each other. The pip daidocs command is a convenience wrapper around the Node engine.

Keys and subscriptions

The reader needs no API key and no subscription. from daidocs import Store only parses .dai files that already exist on disk. It never calls a model and never sends anything to a third party.

Writing new memories (the engine behind the daidocs command) uses each person's own setup, two ways:

• Keyless by default. Through an MCP host such as Claude Code or Claude Desktop, your own assistant does the extraction. This rides on the subscription you already have, and no API key is involved.
• Bring your own key, optionally. daidocs setup offers a provider menu (Anthropic, OpenAI, Gemini, Ollama) if you would rather run extraction against your own account. The key is stored in your user environment only, never written to a file, and you can point at a custom endpoint with OPENAI_BASE_URL.

Either way the key is per person and local to that person. There is no shared or central key.

License

Apache-2.0.