Agent skills
34 packaged skills (20 command + 14 workflow) that route common coding tasks to the right Memtrace MCP tool automatically, installed per-client alongside the MCP server.
What skills are#
Memtrace ships 34 agent skills — 20 command skills (one Memtrace capability each) and 14 workflow skills (multi-step routines that chain several MCP tools toward one outcome). Every skill is a memtrace-*-named markdown file with frontmatter (name, description, allowed tools) that tells your agent when to reach for a Memtrace MCP tool instead of grepping, globbing, or reading files manually.
Skills are a routing layer, not a replacement for the MCP tool reference — each one names the underlying tool(s) it calls, so you can always look up the exact parameters there.
Command skills#
20 skills, each mapped to one focused Memtrace capability.
| Skill | What it does |
|---|---|
memtrace-api-topology | Map API endpoints, outbound HTTP calls, and cross-repo service topology. |
memtrace-cochange | Find files that historically co-change with a target symbol or file, ranked by co-occurrence. |
memtrace-daily | Daily briefing, complexity-x-churn hotspots, and a self-audit session review. |
memtrace-decision-recall | Recall ranked decisions, bans, and conventions from Cortex decision memory by free-text query. |
memtrace-docs-ask | Answer questions from official Memtrace docs only (hosted RAG via ask_docs). |
memtrace-docs-read | Read a full Memtrace docs page by slug (read_doc / memtrace://docs/*). |
memtrace-docs-search | Semantic search over the Memtrace documentation corpus (search_docs). |
memtrace-evolution | Trace source-code change history from Memtrace's symbol-level temporal memory. |
memtrace-fleet-publish-intent | Declare a structural intent before editing in a fleet, so peers on your branch coordinate around you. |
memtrace-fleet-record-episode | Record an edit you just made in a fleet and get its conflict class (A/B/C) against peers. |
memtrace-fleet-resolve | Resolve a Class C fleet decision: submit a verdict, poll your directive, or record a human decision. |
memtrace-graph | Map architecture with PageRank centrality, bridge symbols, Louvain communities, and dependency paths. |
memtrace-impact | Compute the blast radius of modifying a symbol through transitive graph impact. |
memtrace-index | Index a repo into the Memtrace knowledge graph and poll the job to completion. |
memtrace-intent-verification | Verify whether a past decision held or was violated (Held / ViolatedAt / CannotProve). |
memtrace-preflight | One-call pre-flight check before editing a symbol: blast radius, co-change partners, complexity, churn. |
memtrace-provenance | Retrieve the governing decision lineage and contracts that bind a symbol. |
memtrace-quality | Find dead code, complexity hotspots, and refactoring candidates. |
memtrace-relationships | Map callers, callees, references, imports, exports, type usages, and class hierarchy. |
memtrace-search | Hybrid BM25 + semantic search for symbols, functions, classes, types, and logic. |
Workflow skills#
14 skills that chain multiple tools toward a larger outcome.
| Skill | What it does |
|---|---|
memtrace-change-impact-analysis | Compute what a planned change will break — blast radius, affected processes, cross-repo callers — as a risk-rated plan. |
memtrace-code-review | Review GitHub pull requests with the local graph-backed review engine (review_github_pr). |
memtrace-codebase-exploration | Map an indexed repo into a structured overview: scale, communities, central symbols, flows, API surface. |
memtrace-continuous-memory | Keep the index fresh while editing by watching a repo for live, incremental re-indexing. |
memtrace-decision-memory | Umbrella entry point for decision recall, provenance, intent verification, and governing contracts. |
memtrace-docs | Route Memtrace product-documentation questions to search_docs / ask_docs / read_doc before guessing. |
memtrace-episode-replay | Replay one episode’s graph diff — a commit or working-tree save — to inspect exactly what changed. |
memtrace-first | Route code discovery, debugging, and flow-tracing questions to Memtrace tools before grep/glob/manual browsing. |
memtrace-fleet-coordination | Explain conflict classes A/B/C, how Class C gets decided, how to judge, and how branch-scoping isolates fleets. |
memtrace-fleet-first | Coordinate fleets sharing one repo+branch: declare intents, classify episodes, resolve conflicts before they collide. |
memtrace-incident-investigation | Investigate bugs and regressions to root cause using search, impact, call graph, and temporal history. |
memtrace-refactoring-guide | Build a phased, risk-scored refactoring plan from complexity, dead-code, bridge, and impact analysis. |
memtrace-session-continuity | Catch up on everything that changed since your last session using stored session anchors. |
memtrace-style-fingerprint | Pull the codebase's empirical style norm and match it when writing or editing code. |
How they install per client#
Skills are written by the same installer that registers the MCP server, so both land together. Layout is per-client:
| Client | Where skills land |
|---|---|
| Claude Code / Claude Desktop | As the memtrace-skills plugin (~/.claude/plugins/cache/memtrace/memtrace-skills, enabled in settings.json). Local installs (--local) instead write bare copies to <cwd>/.claude/skills. |
| Codex | ~/.agents/skills. |
| Kiro | Auto-inclusion steering docs in .kiro/steering/ (not a skills directory — Kiro reads them as always-on project docs). |
The plugin also ships a references/mcp-parameters.md alongside the 30 SKILL.md files. A standalone shell install path exists too:
$ ./install-skills.sh # Copies skills/commands/*.md (17 files) and skills/workflows/*.md (13 files) # into ~/.claude/skills/<name>/SKILL.md, then prints the MCP registration hint: # claude mcp add memtrace -- memtrace mcp
Or via the npm installer CLI, which writes skills and registers the MCP server together (--skip-mcp writes skills only):
$ npx -y memtrace-skills@latest install # Interactive agent picker when run from a TTY with no --only flag. $ npx -y memtrace-skills@latest install --local # Project-scoped install (used by `memtrace start`'s inline setup hint).
The decision-memory skills (memtrace-decision-recall, memtrace-provenance, memtrace-intent-verification, memtrace-decision-memory) call tools named mcp__memtrace__recall_decision and similar by convention. Those five tools actually live on the separate memcortex-mcp sidecar server, not in memtrace mcp's own registry — see The MCP server for how the two servers relate.
memtrace-first, memtrace-docs, and Rail#
memtrace-first is the routing skill for your indexed source code: it tells an agent to reach for Memtrace graph tools — search, relationships, impact — before falling back to grep, glob, or manual file browsing.
memtrace-docs is the routing skill for official Memtrace product documentation: install, CLI, MCP, fleet, Cortex, enterprise MemDB deploy, skills. It calls the hosted docs MCP tools (search_docs, ask_docs, read_doc) on memtrace.io — see Docs search tools. Use memtrace-docs-ask for Q&A, memtrace-docs-search to find pages, memtrace-docs-read for full page text.
Rail is the enforcement counterpart: a per-client PreToolUse hook (for example .codex/hooks.json, .cursor/hooks.json) that steers an agent toward Memtrace tools at the hook level rather than relying on the model to remember the skill. Skills describe the right behavior; Rail nudges the agent toward it structurally. Enable it with memtrace rail enable (default mode nudge) — see memtrace rail & route for modes, status, and hook wiring.