NORMAL wevibe://landing
verified-knowledge-layer alpha
$ verified_knowledge_layer
The verified-knowledge layer for AI coding

Make your local model
code like a frontier model.

WeVibe is the verified-knowledge layer for AI coding — the fixes that actually worked, with provenance, surfaced the instant they apply. Not memory that stores your slop, but proven knowledge that makes any agent stop repeating bugs your stack already solved — and punch well above its size.

In testing, WeVibe surfaced the right fix as the #1 result 96.7% of the time — and always in the top 5.
get access s see how esc maybe later
~/agent · live session
$opencode run "fix flaky redis timeouts"
✗ Error: connection timeout (ETIMEDOUT)
 node 20.11 · ioredis 5.4 · behind NAT gateway
WeVibe surfaces
verified_fix
verified fix match 0.97
Disable enableOfflineQueue, raise connectTimeout, add keep-alive ping — known NAT idle-drop.
✓ verified by 3 sessions · applies to ioredis 5.x
→ approve & inject? [y]
$ the_problem
The problem

Your agent is confident. It's also wrong a lot.

The fix you need almost always exists — it's just trapped in someone's private chat log, never indexed, never verified. So your agent guesses instead.

err_01

It guesses

It invents fixes that don't apply to your stack or your version — confidently, every time.

err_02

It repeats solved bugs

The whole stack already worked this out. Your agent has no idea, so it makes the mistake again.

err_03

Models leave accuracy on the table

Even frontier models hallucinate. Without the right context at the right moment, every agent — and smaller, local ones most of all — leaves real accuracy behind.

$ how_it_works
How it works

WeVibe knows which past lesson applies — right now.

First-gen memory tools just store everything and match keywords. WeVibe matches the situation — the symptoms, stack, and versions of what you're doing this second — and only ever from fixes proven to work.

step_01 01

It learns when a lesson applies

Every memory stores a situation card — the symptoms, triggers, stack, and environment that make it relevant. Not a bag of keywords.

step_02 02

It matches your moment, apples-to-apples

Your live context and the memory are embedded the same way by the same local model, so the right lesson rises to the top instead of the loudest keyword.

step_03 03

You approve before it touches your agent

Retrieved fixes are decrypted locally, scanned, and shown for your OK. Nothing is silently injected into your context.

situation_card.diff
A situation card, not keywords
Old way — keywords
-
redis, timeout, connection, node
WeVibe — when you'd need it
+
"intermittent Redis timeouts on Node 20 + ioredis 5.x behind a NAT gateway"
$ proof
The proof

From a coin-flip to a near-certainty.

What the verified-knowledge layer actually buys you — measured in our testing against a perfect-knowledge ceiling.

metric_01
96.7%
[██████████████████░░]96.7%
The right fix is the #1 result — and always in the top 5.
Recall@1 · Recall@5 = 100%
metric_02
38→97
old[███████░░░░░░░░░░░░]38
new[██████████████████░]97
Top-result accuracy, old approach → new. A 2.5× lift.
Recall@1: 0.383 → 0.967
metric_03
94%
[███████████████████░]94%
of the perfect-knowledge ceiling captured on answer quality.
+0.69 of +0.73 max lift
metric_04
+0.95
[███████████████████░]+0.95
Weaker models gain the most — a smaller model jumped 1.96 → 2.91 of 3.
nearly matching the ceiling
// During testing, we ran a reproducible benchmark — 60 verified memories across 60 real-world scenarios, with answers graded blind on a 0–3 scale and retrieval embedded locally with nomic-embed-text (768-dim). The memory-lifecycle figures below come from a separate longevity simulation.
$ memory_2.0
Memory 2.0

Not another memory tool. A verified-knowledge layer.

First-gen memory just stores everything your agent does — slop and dead ends included. WeVibe is the layer above it: a curated social graph of fixes proven to work, with the tooling to keep them private, current, and yours. It's the real answer to "why not just ask the model?" — verified provenance, not a confident guess.

proven

Proven, not guessed

Every memory is a fix that actually worked — carrying provenance: who verified it and the exact version it applies to.

private
[#]

Your code never leaves

The network never sees plaintext. Embedding and decryption happen locally at a trusted step — not on someone else's server.

control

You stay in control

Human approval before anything is injected. Self-custodial by design — your keys, your knowledge, yours to keep.

$ the_network
The network gets smarter

Good memories survive. Bad ones fade.

As more verified fixes flow in, the curated graph stays sharp instead of bloating. In a separate longevity simulation, the signal cleanly separated from the noise — automatically.

decay
Good memories that survive94.5%
[█████████████████░] 94.5%
Bad memories that persist15.1%
[███░░░░░░░░░░░░░░░░░░] 15.1%
separation
79.5pt gap

A clean separation between what's worth keeping and what isn't — 5× wider than the prior approach, with good fixes wrongly retired only 5.5% of the time.

// Separate longevity simulation · 9 scenarios × 7 seeds × 300 epochs
$ get_in_early
Get in early

Put verified knowledge
on your stack.

now
Now · Private alpha

Try it on your real work

Bring your stack, point your agent at the pack, and watch the guessing stop.

next

Help build the network

Contribute verified fixes, keep your own keys, and own what you put in — a smarter commons, built on proof.

subscribe
$notify me when WeVibe goes live
// one email, once — the day WeVibe goes live. no spam, no list, no sharing.
request invite j join waitlist g back to top