Built by Recourse Software Inc.Data inside the window
Truncated answers · linear cost · state lost at session end
stateful memory for agentsExternal, stateful memory for AI agents
More efficient AIso you can domore work per token.
Vectorize once. Read at any scale. On any model. One MCP connection turns a terabyte-scale corpus into working memory your agent reads from inside every session that follows.
Recourse Software Inc. — Built in Canada
knowdrive.ai
The problem
Context-window cramming is a dead end.
The industry keeps making the window bigger. That treats storage as a reasoning problem — and the tax compounds three ways.
- 01
The corpus doesn’t fit
A 4,000-document data room hits the ceiling immediately. So does a 20-year client history, or 800,000 lines of legacy code. Crucially important stateful data becomes lossy.
- 02
Cost scales with data, not with thinking
Every query re-ships the same tokens. Frontier-grade prices get paid to hold data in memory — the least intelligent thing a frontier model does.
- 03
Every session starts cold
No persistent state. Users re-paste context daily, and RAG pipelines re-embed the whole corpus the moment they switch models. The knowledge exists — it has nowhere to live.
The insight
Memory belongs outside the model.
Treat the model as the CPU and the corpus as RAM. The drive holds the data; the window holds the reasoning.
Data on the drive, model on top
Complete answers · cost tracks reads · state persists forever
The product
A drive your agent can read — in one connection.
KnowDrive is MCP-native and multimodal. Connect it once and every agent you use — today’s and next year’s — reads the same corpus.
One OAuth MCP connection
Sign in once. Create a store, ingest, and your agent reads it in the same session.
Precise, cited recall
The exact passage, traced back to the unchanged source file it came from, ready to quote.
Multimodal by default
Text, image, audio and video land in the same store and answer the same query.
One corpus, any model
Vectorize once. Switch models, or model versions, and the same corpus answers.
Governance
Multi-tenant and enterprise-authorized from day one, with access control built into the service.
In the product
This is the client we ship, running.
KnowDrive Inspector is a desktop client you can install today. It shows the corpus your agent reads from — what has been ingested, what it cost in atoms, and how sessions hand off to one another.
Every file is tracked from upload through five embedding stages. Atoms are the billable unit, and the count is visible before you are charged for it.
Agent sessions form a real history graph — forks when work runs in parallel, merges when it rejoins. Two people's agents can share one corpus without overwriting each other.
Sample corpora shown. Open the app
Data residency
You choose where the corpus lives.
The same product, provisioned three ways. Pick the one your security review already approves of — the drive is yours either way.
- 01
KnowDrive-managed cloud
DefaultA fully managed, isolated tenant in KnowDrive Cloud. The fastest path to production.
Isolated tenant, KnowDrive Cloud
- 02
Bring your own bucket
S3 · GCS · BlobVectors and source data stay in your bucket. KnowDrive computes over it through a scoped IAM role.
Your cloud account
- 03
On-prem / air-gapped
RestrictedHardened containers shipped to your data centre, with an offline licence and manual updates.
Your data centre
How it works
Four steps. Your agent reads in minutes.
It runs in one session, from an MCP client you already have open.
- Step 01
Connect
One OAuth MCP connection from Claude Code, Claude Desktop, or any MCP client.
~30 seconds - Step 02
Create a store
A card-free personal drive, quota’d and private. Team stores when the team shows up.
1 command - Step 03
Ingest the corpus
Files, folders, URLs, transcripts, media. Vectorized once, versioned, tagged.
same session - Step 04
The agent remembers
Every session after this one starts warm — and the meter only moves when it reads.
activation
The principle
Vectorize once. Read at any scale.
We cut the corpus into small, self-sufficient atoms at embedding time. One fixed write; the reader is chosen later — tiny model or frontier, your call, forever.
The embedding is a one-time write
Cut the corpus once, sized for the smallest reader you might ever want to use. Everything larger works trivially.
The reading is the forever cost
So the reader stays your choice. Switch models, or model versions, and the corpus never moves.
You pay for what you read
Reads scale to zero when your agent is quiet. Storage is a monthly floor, priced per GB.
Interactive · VORAS — Vectorize Once, Read at Any Scale.
One corpus. You choose the reader.
Every dot is one atom of the same embedded corpus. Change the reader and watch the read pattern change shape — the field underneath it never does.
A wide fleet of small readers fans out and takes the corpus in parallel — many tiny reads, everywhere at once.
312 atoms · positions fixed
Readers in parallel
48
Atoms per read
2
Relative read cost
Lowest
Embeddings rewritten0
Switch readers as often as you like — mid-project, mid-conversation, or a model generation from now. Nothing above re-embeds the corpus, because the write already happened, and it only happens once.
Illustrative. The read shapes and counts here are chosen to show the principle — they are not benchmark results or a pricing model.
Shipped
Built, running, dogfooded.
We run Recourse Software on KnowDrive — our GTM brain, memory sync and session handoffs all live in it. The company depends on it to operate.
- Live multi-tenant service with OAuth, billing and quotas
- Every answer carries citations back to the unchanged source
- Satellite skills — memory sync, session handoff — install in one command
- Multimodal ingest: text, image, audio, video
Pricing
We bill atoms. Your bill only grows as your increased knowledge does.
Three meters, and that is the whole model: the atoms we mint from your files, the GB they occupy, and the reads you make against them.
Atomize
one-time, per atom
$1 / 1,000
Store
monthly, per GB
$0.10 / GB-mo
Search
per query
$1 / 10,000
$0
Free
500 atoms, card-free
$19
Launch
/mo minimum
$249
Scale
/mo min · SSO, audit logs
$5,120
Enterprise
/mo min · single-tenant
Usage draws against the monthly minimum before anything is added to it.
Team
Founding Family.
Three brothers. One room. We’ve been arguing about deep systems engineering, enterprise open-source distribution, and consumer-scale product for over 20 years. Now we’re solving it together.
Pierre Augé
Architecture
End-to-end design of the orchestration token optimization layer. Military constrained-systems architect. Patented and deployed systems alongside Canada’s DND/CANSOF.
LinkedIn
Joel Augé
Go-to-market
Built MouseHunt, one of the first Facebook games ever and still running two decades later. Alpha partner on Facebook’s payment, distributed-asset and localization systems.
LinkedIn
Raymond Augé
Platform & community
Ran commercial enterprise open source at scale — managed the code of thousands of developers globally and 600,000+ active deployments.
LinkedIn
Go from more tokens for workto more work per token.
Vectorize once. Read at any scale. Pay for what you read, when you read it, and the meter rests when your agent does.