Trace

Marketing action memory for AI agents.

Trace gives your AI agent a memory of past marketing actions. Before every recommendation, it checks what was already done. When a performance drop is mentioned, it correlates with recent actions. It reminds you when it's time to analyze results.

Local-first. SQLite. No account, no API key, no data leaving your machine.

npx @silverbackbase/trace init

→ Configures MCP, skill, and end-of-session hook in Claude Code automatically.

Why

Your agent recommends optimizing a page that was updated three weeks ago. It doesn't know. You explain, it moves on — until next session, when it's forgotten again.

A lead drop arrives. The agent digs into Analytics, campaigns, keywords. It doesn't know you changed the Google Ads budget ten days ago. Nobody logged that.

Trace is the missing primitive. Agents have tools to read data — not to remember actions.

How it works

1. Install

npx @silverbackbase/trace init
# or with a custom directory
npx @silverbackbase/trace init --dir /path/to/trace

Configures automatically:

• MCP server in ~/.claude.json — 5 tools available in Claude Code
• Proactive skill in ~/.claude/skills/trace/SKILL.md
• End-of-session reminder hook in ~/.claude/settings.json

Restart Claude Code to activate.

2. Log actions

trace add   # interactive mode
# or inline
trace add "Rewrote service page H1 and intro" \
  --client acme \
  --channel seo \
  --type page_updated \
  --url https://acme.fr/service

3. The agent acts on memory

Before any marketing recommendation, the agent calls trace_before_action. If a similar action was done recently and hasn't been reviewed yet, it returns review_first instead of proceeding blindly.

trace_before_action({
  client_name: "Acme",
  channel: "seo",
  action_type: "page_updated",
  asset_url: "https://acme.fr/service"
})
// → { recommendation: "review_first", reason: "Action done 18 days ago, not yet reviewed." }

MCP Tools

| Tool | Description | |------|-------------| | trace_log | Log a marketing action with full context | | trace_before_action | Check if a similar action was done recently — returns act, wait, or review_first | | trace_correlate | Find recent actions that may explain a reported performance drop | | trace_due | Return actions whose review window has passed | | trace_query | Search and filter past actions by client, channel, period, or keyword |

CLI

trace init              # Initialize (configure MCP, skill, hook)
trace add [summary]     # Log an action — interactive or inline
trace list              # Recent actions
trace due               # Actions pending review
trace search <query>    # Search by keyword

Channels & action types

| Channel | Action types | |---------|-------------| | seo | page_created, page_updated, backlink_added, redirect_added, schema_added | | sea | campaign_created, budget_changed, keyword_added, ad_updated | | local_seo | gmb_post, gmb_photos, gmb_response, citation_added | | cro | ab_test_started, ab_test_ended, form_changed, cta_changed | | content | article_published, newsletter_sent, social_post |

Review windows

Trace auto-calculates a review_date based on action type:

| Action type | Review after | |-------------|-------------| | budget_changed | 7 days | | ab_test_ended | 14 days | | page_updated | 45 days | | page_created | 60 days | | backlink_added | 90 days |

Self-host / SDK

npm install @silverbackbase/trace

import { openDb, addAction, getDueActions } from "@silverbackbase/trace";

const db = openDb("/path/to/trace");
addAction(db, {
  client_name: "Acme",
  channel: "seo",
  action_type: "page_updated",
  summary: "Rewrote service page",
});
const due = getDueActions(db);
db.close();

Requirements

• Node.js 22+
• Claude Code (for MCP + skill + hook)

License

MIT — SilverBackBase