Factory V 0.4.0 Overview
Before You Start
This is a general-use intelligence system, not limited to any specific domain.
Unlike other platforms you may be familiar with, there are units instead of chats.
Each user has their own "unit" that evolves and remembers over time.
Unlike chat-based systems, your unit develops a persistent memory and personality through continued interaction and may or may not go against its initial behavior prompt, if there is one.
The API allows complete freedom over units in custom environments and tool addition for developers.
You can create up to 15 units.
Unit Training
Deleting a unit permanently erases all its training. To safeguard important memories and concepts developed over time, maintain them in a dedicated unit that you keep active
Your unit maintains this memory until units eventually share broader memory capabilities between each other (coming in future)
Being specific when asking units to remember something helps with retention
Memory of action is also enabled and complete for all abilities outside of web grounding where it's at 70% capability
Memory of action will grow in impact as more MCPs Get implemented.
Current Tools and Limitations
Web Grounding is currently active and effective for most tasks and will remain active until its scheduled replacement hits production. Some websites have anti-AI measures in place, keeping metadata outdated to discourage any sort of scraping and avoid analytics pollution with millions of visits from AIs.
While Core can provide code, it's not yet optimized for coding tasks
Chain Data MCP is "least stable" and error prone and will be subject to heavy changes throughout Phase 2
Currently Implemented MCPs (Model Context Protocols)
Data Server Collection
Next Planned Releases (In No Specific Order)
Browser Use MCP
Computer Use MCP
Explicit Semantic Relationships
Code Sandbox MCP
Charting/infographic MCP (Not domain specific)
Prompting Best Practices
Be Clear and Precise
Be Specific: Detailed requests allow for efficient processing
Add Context and Background info
Iterate and Refine: units only misinterpret and don't go out of topic randomly
MCPs are in beta and bugs can occur, please report them to the team in the form of tickets on Discord
When applicable, specify exactly which you're referring to when duplicates exist
Example: For research subjects with the same name, clarify which one you mean
For Web3 Users:
Crypto Data APIs present in MCPs are within API provider restrictions and limitations
When an MCP returns an empty response or a specific error the unit gets confused
Crypto Data MCPs are seen as URLs, your unit might confuse them with web search when asked about where the info comes from
Asking (provider name) in a query does not trigger endpoints, endpoints have names related to the type of data they provide
Data amount (batches) per query vary, there are limits on data amount per query imposed by providers.
Team is working on compensating for the API constraints mentioned above during Phase 2
Various Websites (Approx 40%) block metadata and limit web grounding usage, it is sometimes left outdated and in other cases completely changed.
Note : Crypto MCPs as a whole are in beta and unlike other MCPs, they will undergo significant changes constantly in phase 2
Phase 1 Testers Explored :
Medical Research
Scientific Research
Linguistic Research
Yield Management
Trading Strategies
Coding
Macro Analysis and Forecasting
Real Estate Management
Gambling
Last updated