Space

01 / Brief

Solution

Space is a workspace for persistent AI agent memory, built to reduce the context lost between conversations and tools. It brings documents, tasks, structured data, and saved preferences and decisions into one place that authorized agents can access through MCP. The web interface lets users inspect the same work, follow memory back to its sources, correct it, and manage agent access. The engineering combines a TanStack Start web app, shared FastAPI commands for MCP and REST running on Cloudflare Containers, PostgreSQL with pgvector for retrieval, and encrypted content revisions. Space is in active development; public access is not open yet, and integration availability varies by agent.

02 / Constraints

Problem

• Retrieving useful context without carrying an entire conversation history into each new session
• Keeping agent writes and manual website edits consistent through shared commands and revision checks
• Balancing retrieval speed and evidence quality while preserving source references
• Protecting stored content while enforcing permissions on every agent request

03 / System

Architecture / workflow

• MCP and REST access share the same authorization rules, workspace commands, and saved revisions
• Documents, tasks, and structured data stay alongside persistent memory and its source references
• Semantic retrieval uses PostgreSQL and pgvector to find relevant context across sessions
• Jev supports evidence assessment and memory classification into categories and projects
• Encrypted content revisions, scoped agent permissions, and revocable connections keep access explicit