nodecode-recall

Recall: a private, on-device index of the documents on your machine, so "where is that Acme invoice" is answered from the file instead of a guess. A Nodecode rewrite of memeshee/hermes-recall by memeshee.

Get it

Tell the agent "add the recall cell", or run nodecode add nc://recall.

Set it up

Nothing is required. The index is ~/Documents alone until you name more:

"recall": {
  "roots": ["~/Documents", "~/Downloads"],
  "excludes": ["*.zip"],
  "ambient": true
}

~/Downloads and ~/Desktop are opt-in on purpose: they are full of screenshots, installers and exported keys. Then say "index my documents", or run /recall index.

A PDF is read through pdftotext and a Word, Excel, PowerPoint or OpenDocument file through unzip, when your machine has them; without either, the file is still indexed by the words in its name and recall_open says what it could not read.

Use it

The model gets five tools, and the same verbs are callable from eval:

Tool	What it does
`recall_index`	Builds or refreshes the index over the roots (or the folders the call names). Incremental: only new or changed files are re-extracted. `rebuild` re-reads everything.
`recall_search`	Ranks the files under a query, with the matched line in brackets, a date and a score. Filter by `path` substring or by `ext`.
`recall_open`	One file as text, with `offset` and `max_chars` to page through a long one.
`recall_stats`	What is indexed, where, when, and what was skipped.
`recall_forget`	Drops a file, a folder or the whole index. It never touches the file on disk.

/recall <query> searches, /recall index [path] builds, /recall stats and /recall forget all do what they say. (help :recall) is the manual, and (recall:search "powder coating") works from eval too.

With ambient on, a message that reads like it is about a document ("send the lease to legal") gets a short block before the model answers: up to max_hints paths, each with a date and one bracketed line, and never the file itself -- the model still has to read it before quoting it. A file already shown to that session is not shown again.

What it never reads

One policy decides, in guard.lisp, judged on the resolved path so a symlink named notes.txt cannot point at a key:

Nodecode's own home, where its cells, secrets and store live;
virtual filesystems (/proc, /sys, /dev, /run/user) -- /proc/self/environ is a readable file full of keys;
the credential-store denylist: .env, *.pem, id_rsa*, *.kdbx, .netrc, *.gpg, *secret*, *recovery*code*, .aws, .ssh, .gnupg, .kube, .config and the rest of it in guard.lisp.

The rules are the cell's: excludes adds patterns, it never removes one. Nothing leaves the machine -- there is no network call anywhere in the cell, and no embedding service.

What is in it

File	What it is
`package.lisp`	The package, the index structures, the one index file under the organism's home.
`guard.lisp`	The read policy: the home, the virtual filesystems, the credential denylist.
`extract.lisp`	Text, HTML/XML, PDF and Office extraction, and the chunker.
`index.lisp`	The walk, the incremental run, the postings, BM25 and the snippet.
`cell.lisp`	The five tools, `/recall`, the manual, the ambient hint and the background refresh.

nc://recall