“Why does it keep forgetting?”
Useful work spans days, files and decisions. Chat windows are not enough for long-running project memory.
Non-commercial local AI infrastructure project
AI is powerful. The way we work with it is still broken.
Most people know AI as ChatGPT, Gemini or Copilot: a prompt box, a smart answer, and then a lot of manual cleanup.
AIWorkspace is being built as the missing workbench around AI: memory, files, internet research, evidence, task state, local models and human control.
What breaks after the impressive answer?
People do not just need AI answers. They need AI work they can trust.
The first demo is magical. The tenth real task is different: context disappears, files drift, sources are unclear, actions may repeat, and nobody can easily explain what happened later.
“Where did that come from?”
Serious work needs sources, state and retained evidence, not just confident prose.
“Can it research safely?”
AI needs current information, but internet-enabled agents should not become uncontrolled host-level authority.
“Does it understand my actual work?”
Documents, images, code and notes need to become part of a governed workflow, not scattered attachments.
“What exactly did it do?”
When AI can act, people need boundaries, replay protection and reviewable outcomes.
“Private, but raw.”
Local models are promising, but without a work layer they are often just isolated chatboxes.
The simple comparison
Cloud AI is convenient. Raw local AI is private. AIWorkspace is building the governed workbench.
| Need | ChatGPT / Gemini / Copilot | Raw local model | Broad local agent | AIWorkspace direction |
|---|---|---|---|---|
| Strong answers | Excellent | Model-dependent | Model-dependent | Uses available models |
| Local control | Limited / cloud-shaped | Strong | Strong | Core principle |
| Internet research | Often available | Manual or add-on | Powerful | Governed research lane |
| Project memory | Platform-dependent | Usually weak/manual | Varies | Structured and inspectable |
| Evidence trail | Limited | Weak | Varies | Designed into workflow |
| Action safety | Platform-controlled | Manual | Can be too permissive | Boundaries and review |
| Institutional fit | Useful but opaque | Private but raw | Often hard to govern | Built around governance |
Internet, with a steering wheel
Local AI should not mean isolated AI.
AIWorkspace is being built so AI-assisted work can use the internet where useful — research, public sources, documentation, examples and current information — while keeping the workflow visible, bounded and reviewable.
The goal is not to give an agent unlimited control over a machine. The goal is useful internet-assisted work with evidence, boundaries and human oversight.
Visual bars are conceptual design goals, not measured benchmark scores.
The workbench
The model is the engine. AIWorkspace is the workshop.
Known AI
ChatGPT, Gemini, Copilot, Claude, Qwen and other models
→
AIWorkspace layer
Memory · Files · Internet · Evidence · Tasks · Boundaries · Review
→
Serious work
Remembered, checked, recovered, reviewed and kept under local human control
What already works
Early, not finished — but no longer just an idea.
AIWorkspace has reached a verified local baseline. Publicly, that means six practical behaviours have already been checked without publishing the internal recipe.
Reports healthy local runtime state.
Handles text follow-up context.
Can continue image-related discussion after upload.
Journals side effects with duplicate/replay protection.
Uses governed model-backed local execution.
Can verify architecture boundaries with zero current boundary failures.
What comes next
A flood of local AI possibilities — but governed.
The ambition is broader than chat: local research, coding, documents, image workflows, playable HTML demos, generated assets, long-running project memory and benchmarked comparisons against raw local-model workflows.
Internet research
Local coding workflows
Image upload and review
HTML game demos
Documents and reports
Text-to-image lanes
Benchmark arena
Evidence packs
Benchmark arena — planned, not claimed
Next: prove the work advantage with real comparisons.
AIWorkspace is being prepared for direct comparisons: cloud chatbot workflow vs raw local model workflow vs governed local AIWorkspace workflow. Planned measurement areas include task completion, context retention, repeatability, recovery, evidence quality, local control and operator review.
No fabricated performance numbers are claimed here. When benchmarks are published, they should be reproducible and source-backed.
Public-interest review
Non-commercial. Local-first. Built for serious review.
AIWorkspace is currently a non-commercial research and infrastructure project. It is not offered as a paid product or commercial service. The purpose of this website is to support technical review, public-interest discussion and responsible development of local AI work infrastructure.
A redacted evidence pack is available for qualified technical review under appropriate conditions.
Sources used for public problem framing
Statistics Canada: workplace generative AI use rose from 17% in September 2024 to 30% in July 2025. KPMG Canada: 51% of Canadian employees used GenAI at work in 2025. TELUS Digital: 68% of enterprise employees using public GenAI used personal accounts and 57% entered sensitive or high-risk information. Pew Research Center: roughly half of adults in several surveyed countries said they are more concerned than excited about AI in daily life.
Statistics Canada · KPMG Canada · TELUS Digital · Pew Research Center