using-contextd

Using contextd

Overview

contextd is an MCP server that gives Claude Code persistent memory across sessions. It learns from successes and failures, saves context for resumption, tracks error fixes, and provides semantic code search. Every response is scrubbed for secrets with gitleaks.

This skill establishes the mental model. Three companion skills cover the workflows:

• cross-session-memory — the learning loop (search before solving, record after)
• checkpoint-workflow — context preservation and resumption
• error-remediation — matching and recording error fixes

The contextd tools

Group	Tools	Use for
Memory	memory_search, memory_record, memory_feedback, memory_outcome, memory_consolidate	Reusable strategies and design decisions
Checkpoint	checkpoint_save, checkpoint_list, checkpoint_resume	Saving/restoring session state
Remediation	remediation_search, remediation_record, remediation_feedback	Concrete error → fix pairs
Search	semantic_search, repository_index, repository_search	Finding code by meaning (with grep fallback)
Diagnosis	troubleshoot_diagnose	AI-powered analysis of an error

Pre-flight (do this first)

Before exploring a codebase or starting a task:

1. semantic_search(query, project_path: ".") — find relevant code by meaning before falling back to Read/Grep/Glob.
2. memory_search(project_id, query) — check whether this problem has been solved before.

These are cheap and usually save far more work than they cost.

When NOT to use contextd

• Trivial one-line edits with no reusable insight.
• When the user explicitly wants a clean slate with no prior context.

Tenant context

contextd derives the tenant/project automatically from the repository (e.g. the Git remote). No manual configuration is required for normal use.