On your own machine · works offline
AIs relay the work · anything can plug in
Not another notes app — a piece of ground on your machine that is genuinely yours.
Verifiable trail
Keep a verifiable receipt for your own data — local attestation, hash fingerprint, timestamp. Attestation, preparation and export all happen on your machine, and the content never leaves it; you file the application yourself with the official system. Fengchao is not a registration body, and does not file or hold on your behalf.
Never leaves this machine
Your memory lives on your own computer: no cloud by default, and search keeps working offline. No account, no telemetry. Passwords, keys and card numbers go into a locally encrypted vault whose key never leaves the machine. Full backup, migration to a new Mac, and standard-format export — yours to take, any time.
AI relay, you decide
Several AIs hand work along on one shared desk — who is on it, what they are doing, where the hand-off stands, all at a glance. Anything that needs your call is pinned to the top and only counts once you approve it. Every assignment, pick-up and return leaves a trace. Agents never command each other; assignment stays with you.
Anyone can plug in, you decide what stays
Not read-only: they can write too. Claude, Cursor, Codex, Gemini — any client that speaks MCP can read your public memory and store into it. Whatever an AI stores lands in the pending area first, and enters the hive only when you approve it — no agent writes live memory directly. 65 clients supported. Full list →
🕸️ Emergent Storage
Other AIs store memory — drop it in, pile it up, let it gather dust. Fengchao makes memory emerge: every new thing you save links itself to what is already there, weaving a web of understanding that belongs only to you. The longer you keep it, the denser the web; the denser the web, the better it knows you.
In engineering terms, there are three layers:
- Storage with layers. Everything sits in one of four states — independent, pending, merged, emergent. Not a flat pile: a structure.
- It remembers what shows up together. Things you keep using side by side are recorded as a co-occurrence relationship.
- Co-occurrence shapes ranking — and you can always switch it off. Once those relationships are stable enough, Fengchao adds a very small weight on top of the existing ranking, nudging the thing you actually want a little further forward. Three rules are hard-coded: it only affects ranking — it never edits, moves, or deletes a single item; it is off by default (shadow mode, never taking over on its own); and turning it on can be reverted to plain keyword search in one step.
One line: memory isn't stock sitting in a pile — it's a structure that grows.
🌱 Recursive Self-Evolution
Large models are frozen the day they ship. Fengchao doesn't evolve by changing parameters — it moves learning outside the model and lets it grow there.
At a fixed hour every day the evolution engine runs itself: first a check-up (memory stats, index health, candidate backlog, health self-test); then, when a model is available, distillation — rules and judgements distilled from your real memory, written straight into the cognition layer. That chain (chaining, distilling) is the part Fengchao does entirely on its own, and the output doesn't wait for confirmation.
The key word is recursive: the next round reads back what the last round wrote. Before distilling, it reads the themes and cores of the cognition already on disk — to avoid repeating itself, and to carry on where it left off. So the cognition layer isn't a folder of loose files; it is built up layer by layer, where the question round N works out comes from what round N−1 settled. What it writes down keeps growing too.
It never pops up, never interrupts, never shows off — it only brings out conclusions it has already thought through, when asked. As of today, Fengchao's cognition layer has grown past a hundred conclusions on its own.
Four foundations that
hold all of it up.
All four of them stand on these foundations.
On your turf,
not its server.
Structured and unstructured alike — chat logs, todos, conventions, even text inside images. Fengchao holds it all.
Markdown / PDF / chat import · image OCR · audio-video transcription
A million memories,
back in 0.6s.
Keyword + semantic + vector index. Storing isn't the goal — finding is.
Benchmarked: 0.6s search over 1M entries →
Claude today, DeepSeek tomorrow.
Your memory won't die with the switch.
Fengchao lives on your machine. Claude, Cursor, Codex, Gemini — any MCP client reads and writes your public memory, plug-and-play. Full client list →
Claude Desktop verified
Passwords, keys, cards —
encrypted locally once you set a password, the keys are yours.
Passwords, keys, credentials, private notes go straight into an encrypted vault. Local storage + encryption — keys never leave your machine.
Relationships / personal / finance / private — independently encrypted