daidocs

daidocs command [options]daidocs setup [options]daidocs convert [options]daidocs pending [--route project] [--to folder]daidocs stores [--scan folder] [--forget id]daidocs backup folder --to destination [--zip] [--memory-only]daidocs scrub [store] [--apply]daidocs ingest folder [store] [--observer spec] [--force]daidocs ask [store] question [--actor spec]

daidocs is the command-line front end for DaiDocs, an Apache-2.0 plain-text format for assistant memory. Each conversation becomes a .dai file with three UTF-8 zones (YAML header, fenced JSON, and text) plus a small derived index. The files live on your disk and can be read with grep, git, or any editor.The npm package also installs daidocs-setup, daidocs-server (stdio MCP), and session hook binaries. npx daidocs setup is the documented one-shot install. A Python extra (pip install daidocs) exposes the same daidocs command and a Store reader; the engine itself still needs Node.js 18+.The default shared store is ~/DaiDocs. Declared project folders keep memory in .daidocs/store inside the folder. Session hooks (when installed) convert in the background after about 4,000 new tokens; shorter sessions wait in _unconverted/ and are still injected into the next session verbatim.ingest and ask are the only subcommands that call a model themselves. Conversion through Claude Code hooks is done by the assistant already in the session and needs no extra API key.

setup

Detect assistants on this machine (Claude Desktop, Claude Code, Cursor, Windsurf, Codex, Cline, Continue, Zed) and write MCP / hook / icon / reading-protocol config, backing up every file it touches.

With setup, print each switch, whether it is on, and the command that changes it.

With setup, choose each surface interactively instead of configuring all of them.

Put back files setup changed, or remove what was installed.

Model that writes .dai files (for example openai:gpt-4.1-mini). Applies to conversion, hooks, MCP, and the CLI until changed.

Declare the current folder as a project, set its type (normal, locked, frozen, connected, shared, confidential, temporary), and choose a per-folder store or the shared ~/DaiDocs store.

Turn chat history into .dai files. Interactive by default. Reads Claude Code sessions under ~/.claude/projects, _pending / _unconverted backlog, or a folder of .txt / .md / .jsonl.

Limit convert to Claude Code history, already-captured unconverted text, or an export folder.

Choose which items, where to write them, and skip prompts.

List captured-but-unconverted sessions, grouped by project.

Convert one project's pending group into that project's store.

List declared stores and any whose folders have gone missing. Ids live in ~/.daidocs/stores.json.

Copy a project to a dated destination, skipping node_modules and .git. --zip writes a zip; --memory-only copies only .daidocs.

Report credentials already stored in .dai files. --apply rewrites them out.

Convert a folder of documents into a store. Calls the observer model (can cost API usage). --force re-ingests files that are already present.

Answer a question from a store on the command line. --actor spec selects the answering model (can cost API usage).

DAIDOCS_STORE

Default store path. Default: ~/DaiDocs. When set, the store argument on ingest, ask, and scrub can be omitted.

Override the observer model for one run.

Skip the auto-archive hook.

Decline the optional one-time anonymous install ping without being asked.

.dai files are not encrypted: confidential only stops other stores from reading them, it does not hide them on disk. scrub is not a guarantee that secrets are gone. Agent traces (tool calls, stack dumps) convert poorly. Below roughly 20k tokens of history the store adds little over keeping the raw transcript. ingest and ask spend tokens on a paid model when you are not on a subscription path that writes the extraction itself.

DaiDocs launched in September 2026 as an open format plus a Node reference engine and MCP server, maintained by Kerneta under Apache-2.0.

claude(1), grep(1), git(1)

Source code · Homepage · Documentation