The MCP that knows who you are before you talk

The problem with any LLM assistant is context reset. I start a new conversation and the model doesn’t know I run a small packaging company in southern Spain, that my mix is 80% B2B, that I prefer direct recommendations, that my brand uses Kraft #C9A36B and not bright red. Either I retell it every time, or paste a brief, or settle for generic answers.

The solution I adopted: a Personal Context Portfolio deployed as an MCP server on Railway. Ten Markdown files — identity, communication style, preferences, goals, active projects, tooling, role, team, domain knowledge, decision log — accessible via SSE from any Claude conversation.

Any Claude instance (Cowork, claude.ai, API) can call get_context("identity") and start a session already knowing who I am. More importantly: it can call update_context() or append_to_decision_log() when we learn something new. The portfolio stays alive without me having to edit anything.

Context is infrastructure, not a brief you rewrite every Monday.

Decisions I made when building it:

• Markdown over JSON. I read it, the model reads it. One format.
• Modular, not monolithic. Ten files by domain instead of one giant blob. The model loads only what’s relevant.
• Decision log as append-only. Every significant decision gets logged with date and reasoning. The next agent understands how I think, not just what I decided.
• Railway over AWS. Deploy in 10 minutes, $5/month, zero ops.

Two months in. The operational difference: I went from explaining who I am in every chat (1–2 paragraphs) to talking about the problem directly. It’s a small friction saving per conversation, but it compounds when you work with agents daily.