Build an AI Dev Memory System for Client Work

Switching between client projects means rebuilding mental context every time you open a session. The architectural decisions from last week's call are gone. The API quirks you discovered on Tuesday vanish by Thursday.

This is worse for AI-assisted development because your AI assistant starts completely fresh each session. You end up re-explaining project constraints, tech stack choices, and client preferences before any real work happens.

The cost compounds across clients. Three active projects means three separate context rebuilds per day, each burning 10-15 minutes of expensive AI conversation time on orientation instead of output.

The simplest isolation pattern is one vault folder per client project. Each vault contains only that client's decisions, patterns, and references. No cross-contamination, no accidental leakage of one client's architecture into another's codebase.

Tag entries with the project name, current phase, and decision type. A typical client vault might use kinds like decision, pattern, and reference with tags for the specific domain area. This structure scales cleanly from a solo freelancer to a small team.

Keep client data isolated by default. When you find a generalizable pattern that applies across projects, save it to a personal vault with a generic tag rather than duplicating it across client vaults.

During a work session, save architectural decisions, API quirks, client preferences, and implementation patterns as they come up. Use save_context with specific titles and relevant tags so future retrieval is precise.

Before starting the next session, retrieve recent entries to prime your AI with relevant context. A single get_context call with the project name returns the most relevant decisions and patterns, weighted by recency.

This creates a virtuous cycle. Each session deposits context that makes the next session faster. After two weeks, session startup drops from 15 minutes of re-explanation to under two minutes of automated context loading.

For three to five active projects, use a consistent folder convention. Each client vault lives in its own directory with the same internal structure: decisions, patterns, references, and session notes organized by kind.

Use kinds to separate content types cleanly. Decisions capture why something was chosen. Patterns capture reusable approaches. References capture external docs, API endpoints, and credentials locations. This taxonomy works across any client project.

When cross-pollinating solutions between clients, search your personal vault for generic patterns rather than searching client vaults directly. This keeps boundaries clean while still benefiting from accumulated experience across all your work.