One local knowledge graph feeding Claude, GPT, Gemini — or a small model on your own machine. Stop pasting context into a blank box. It lives once, persists across every model, and compounds with every session you finish.
The most capable AI tools were built for people who want to operate AI — endless settings, prompts, and plumbing to master before you produce anything. Waggle takes the opposite bet: your time is too valuable to spend on AI plumbing, so we run it underneath and hand back finished work. You didn't become an expert to babysit a model.
You're known here — never re-introduce yourself, never start from a stranger. It already knows how you think and how you like your work done.
Nothing is lost. Walk back into any project after a day, a week, a quarter — it's all still there, exactly as you left it.
You stay the expert. The prompting, routing, and scaffolding run underneath — on tap when you want them, out of your way when you don't.
Spreadsheets to the last detail, decks that hold up in the room, docs polished to the final word. Everything you ship is genuinely your best, and unmistakably yours.
Local-first and model-agnostic — runs on any model, even a small one on your own machine. Yours to keep, export, or delete, anytime.
Want the prompts, the routing, the raw model picker back? They've never gone anywhere — Waggle just stops making you use them.
You bring the judgment. Waggle brings the AI. Open it once and your context and your work-in-progress are already there — nothing re-explained, nothing lost.
Start fresh and it learns you fast — or import your history. One-click import from ChatGPT, Claude, and Cursor: brought in on your machine, stays on your machine.
Open it and you're already moving: here's where you left off, here's what ran overnight, here's what's next. Your work picks up mid-stride, never from a blank box.
You make the calls; Waggle runs the AI underneath and hands you finished work. Close it, and all of it stays right where you left it for next time.
Launch Claude Code, Cursor, or Codex straight from a Waggle workspace and it opens already knowing the work — then commits what it does back to the hive. No copy-paste, no re-explaining. Nothing leaves your machine.
The agent recalls this workspace's decisions, constraints and history the moment it starts — no prompt-stuffing.
Everything it does flows into the hive with provenance, so Waggle and your next session pick up where it left off.
Hooks install and uninstall cleanly. The memory stays on your machine; you can verify or remove anytime.
On LoCoMo — the field's standard test for long-term conversational memory — Waggle's open-source substrate scores 87.66%, a new state of the art. Measured under the prior leader's own protocol and judge, whose pipeline we reproduced to within 0.03 points before comparing.
That's ~1 point off the full-context ceiling — what a model scores with the entire conversation pasted in. The memory recalls almost as well as having everything in front of it.
Every memory operation runs on-device — SQLite, local embeddings, an in-process reranker. Warm recalls in 58–83 ms. The strongest LoCoMo result on record for a fully local system.
The substrate and the benchmark harness are public. Fork it, read the graph, reproduce the score offline. No black box — github.com/marolinik/hive-mind.
LoCoMo, N = 1,540 questions, GPT-4.1-mini as answerer and judge — the prior SOTA's exact published protocol, reproduced in-harness to 0.03 points before comparison. The intelligence lives in the memory layer, not the model, so it travels with you — even onto a small model on your own machine. Honest caveat: we spend more tokens per question than the leanest systems; accuracy first, efficiency work is underway.
Your skills aren't static prompts. Waggle re-optimizes them against real outcomes, keeps every version with the score that earned it — and the good ones spread across your agents and workspaces. The longer you use it, the better it works.
Each skill is re-tuned against held-out tasks and judged by an ensemble of models. A new version only ships if it actually wins — and the old one stays, so you can always roll back.
When one agent learns something good, the others don't start from zero. Skills and tool setups diffuse across your agents and workspaces — the waggle dance. On a team, the whole org levels up at once.
Hunters chase, builders ship, orchestrators coordinate. Waggle remembers — across roles, across tools, across the moments in between.
Three tiers, no feature-count games. You pay for the scale of the team using the memory — not for arbitrary check-marks. Memory is free forever.
Everything Waggle does — on your infrastructure, your permissions, full audit trail. Your data never leaves your perimeter. That's KVARK, by Egzakta Group.
Bring your work home to the one place that knows you, knows your projects, and makes everything you ship genuinely your best — then pick up exactly where you left off.