dumpgenius

What it is

dumpgenius is your single front door for dumping anything — notes, ideas, half-thoughts, photos, files, links, voice, "shit you're not even thinking about." You throw it in from wherever you are; it lands losslessly and forever, then quietly gets useful: classified, described, searchable, and routed onward.

One sacred rule: never lose a dump. Capture happens first and is irreversible-safe. Everything after — classify, extract, search, ship, back up — is derived and can always be re-run.

---

Status at a glance

Feature	State	What it means for you
Telegram capture	✅ Live	Send/forward anything to the bot from your phone.
Save from openclaw	✅ Live	Tell your openclaw assistant "save / keep / remember this" and it lands in the lake. openclaw is the talking front door; dump is its durable memory.
Lossless lake	✅ Live	Every dump is saved raw, immutable, deduped. Nothing is ever lost.
Classify + tag	✅ Live	Each dump gets a class (note / idea / receipt / contact …), a summary, tags.
Photo & vision extraction	✅ Live	Photos get a rich description + pulled-out data (business cards, labels, serials, prices).
Semantic search (recall)	✅ Live	Ask "what did I dump about X?" and get it back, even from a photo.
Off-device backup	✅ Live	Everything is copied to the NAS automatically, every cycle.
Private by default	✅ Live	Processing runs through local models — nothing has to touch a commercial API.
Routing rules	✅ Live	Declarative rules can ship/archive/back up by class (local targets for now).
Logging	✅ Live	Leveled, greppable logs for every surface.
Web inbox UI	🔜 Next	A simple page to browse + search your dumps in a timeline.
More connectors	🔜 Next	Email-forward, Google Drive watch folder, paste-into-UI, API.

---

How to use it

📱 From your phone — Telegram

Send or forward anything to the bot. It lands instantly and replies with a confirmation.

• Bot: @DumpGeniusBot
• Works with: text, notes, links, photos, documents (PDF, Word, spreadsheets, code, archives…), voice notes, audio, video, and forwarded messages (the original source is kept).
• Commands: /help shows what it accepts. Everything else you send is treated as a dump.
• Limit: 20 MB per file (a Telegram rule). Bigger files go via the command line.

⌨️ From the command line

The dump tool is the capture + query spine. Common moves:

dump "a raw thought"          # capture text
echo "..." | dump             # capture piped stdin
dump -f ~/file.pdf            # capture a file (any type)
dump -c                       # capture the macOS clipboard
dump ingest ./folder          # capture every file in a folder

dump ls                       # list recent dumps
dump show <id>                # show one dump (metadata + extracted fields + raw)
dump stats                    # counts, bytes, where the lake lives
dump recall "that idea about pricing"   # semantic search across everything

dump orient                   # classify + embed any un-processed dumps
dump route                    # preview routing rules (add --apply to run)

Private mode (nothing leaves the machine — local models): prefix with DUMPGENIUS_PRIVATE=1.

🤖 From openclaw

openclaw is your conversational assistant; dump is its durable "save this" memory. Just tell openclaw in plain language:

• "save this / keep this / remember this" → openclaw lands it in the lake (text, a photo,

a file) and confirms with the saved id.

• "what did I save / dump about X?" → openclaw recalls it from the lake.

Conversations stay in openclaw; only what you ask to keep becomes a durable dump. (Phase 1 is explicit — you ask. Letting openclaw decide what's worth keeping from context is next.)

---

Roadmap — what's next

In rough order:

• Web inbox — a clean page to scroll your recent dumps, see their class + extracted data, and search. (The search engine already exists; this is the friendly front-end.)
• Deterministic workflows — small, reliable per-type flows: resize a photo for the web, turn a business-card photo into a contact, turn notes into a shareable report. Hands the system-access steps to openclaw; keeps the predictable logic here.
• More front doors — email-forward (forward a mail and it lands), Google Drive watch folder, a paste-into-UI box, and an API endpoint for other tools/agents to push dumps.
• Always-on extraction — move the vision/classify model onto an always-on host so photos get processed instantly instead of next-time-the-laptop's-awake.
• Cloud mirror + a third backup copy — optional off-site sync and a parity-protected archive.

---

Under the hood (the map)

Names only — no addresses, no logins. The flow:

Your phone / terminal / openclaw → apollo (always-on capture host; lands the raw
bytes the instant they arrive) → the lake (immutable, content-addressed store + index,
on local SSD) → the model host (a Mac on the private network; classifies + describes
over the LAN when it's awake) → UGREEN NAS (automatic off-device backup every cycle).

• openclaw feeds the lake — openclaw runs on its own Mac on the private network; when you ask it to save something, it hands the bytes to apollo's capture over the tailnet, stamped as coming from openclaw. The lake is the single durable memory behind it.
• Capture is model-free — it never waits on anything. If the model host is asleep, dumps queue and process on the next pass. Capture is never blocked.
• Private by default — orientation runs on local open-source models; commercial APIs are opt-in, not required.
• Code lives in a private repo (mmbianco78/mb-dumpgenius); deploys are a git pull + service reload.
• Isolation — the capture service runs at low priority and opens no inbound ports, so it never disturbs anything else on the host.

---

The one promise

Throw it at dumpgenius and forget it. It landed. It's safe. It's findable. That's the whole point.