Case 01 · Infrastructure

Memory across tools.

Open Brain is the layer that keeps three of my AI agents in sync, across six applications, without context loss. Built on MCP in January 2026, before the protocol was widely adopted. Running in continuous production since.

Role
Architect and builder
Duration
Two days to first working version, six months of evolution
Team
Solo build
Vertical
AI infrastructure
Annotated architecture diagram of Open Brain. Three agents (Claude, Crafty, Hermy) on top, MCP and REST API in the middle, PostgreSQL with pgvector at the bottom, plus a Voyage AI embeddings layer and cron processes for weekly consolidation and monthly decay.

The problem

If you work with more than one AI tool, you have noticed they do not share context. Whatever you told Claude on Monday is invisible to ChatGPT on Tuesday. Each vendor builds memory inside its own interface. None bridges across, which is where the friction lives in a multi-agent workflow.

I started Open Brain in January 2026 while MCP was still early. The bet was that an open layer beats a vendor feature, and that the protocol was about to become standard. Six months later the layer is still running, and the architecture has held up.

Three things vendor memory keeps getting wrong

  1. Lock-in. Anthropic remembers Claude, OpenAI remembers ChatGPT. Neither will bridge to the other.
  2. No taxonomy. Pure semantic search over raw thoughts gets noisy fast. Without typed structure (project_update, decision, pattern, insight, action_item, context) retrieval becomes a wall of vaguely related text.
  3. No decay. Everything weighs the same forever. A throwaway note from March ranks next to a critical decision from May. Without a half-life, memory suffocates itself.

What I built

A layer between any AI tool and a structured knowledge graph. Two entry points: an MCP connector for protocol-aware tools (Claude Desktop, Cursor, CoWork) and a REST API for the rest. Storage on PostgreSQL with pgvector, embeddings via Voyage AI, deployed on Railway. Retrieval is task-aware, with planning, reflection and briefing modes. A monthly decay process archives stale thoughts. A weekly consolidation clusters related ones into topic patterns.

Seven MCP tools expose the system. Three agents read from it daily. Claude is my primary partner, Crafty runs skills on my Mac, Hermy handles autonomous background work.

Cost discipline sits in the design rather than retrofitted. Embeddings are batched and cached. Retrieval stages through a summary-first pass before expanding into full bodies, which keeps token use low on the common case. Token-aware routing across the read and write paths.

Flow diagram: a thought captured in Claude travels through the MCP layer and becomes available to Crafty on the Mac and to Hermy as a background process.
One shared layer. Many tools on top. The memory belongs to the user, not to the vendor.

Six months in

The architecture is the same one I would put inside an organisation. Storage tier, embedding layer, retrieval modes, decay process, MCP and REST endpoints. What changes between personal and organisational use is the data sources, the governance around them, and the audit trail. Not the underlying design.

What this means for you

If your organisation uses more than one AI tool, you already have the problem this layer solves.