Every AI coding session starts from zero. Retrace gives your agent persistent memory — it logs decisions, bugs, and context as you code. Stored locally in SQLite. No account needed.
Install in 10 seconds. Sign up when you want more.
// Agent logs rich context automatically
→ log_entry({
content: "Fixed Redis TTL causing stale responses in Forbes iOS app.
Root cause: 24h TTL but data changes hourly",
category: "debugging",
project: "forbes-mobile",
people: ["Marcus", "Sarah"],
refs: ["MOB-1634", "PR #287"],
component: "cache-layer"
})
✓ Logged entry #142 at 2026-07-22 14:30
// Months later, recall naturally
→ recall({ query: "caching problem in the mobile app" })
✓ Found 3 entries — showing top result:
[#142] Jul 22 [debugging] (forbes-mobile) [cache-layer]:
Fixed Redis TTL causing stale responses...
People: Marcus, Sarah | Refs: MOB-1634, PR #287
Your data stays on your machine. No account, no cloud, no data leaving your laptop. Upgrade when you're ready.
SQLite on your machine. Zero config — just run npx and start coding. Your journal lives at ~/.retrace/journal.db. You own it completely.
One command: npx retrace-mcp@latest. No account required. No API keys. Works offline. Your agent is logging in under 30 seconds.
Your AI agent logs entries for you. Decisions, bugs, people, references — captured automatically as you work. You never open an app to write.
Ask "What was that caching problem last month?" — finds entries by content, people, project, and component. Works even when details are fuzzy.
"Who knows about the payment integration?" — finds team members with relevant experience based on their logged involvement.
GitHub Copilot, Cursor, Claude Desktop, Zed — any MCP-compatible agent. Best with agent instruction files that tell it to log proactively.
Install, code normally, query anytime.
Run one command. That's it — no account, no config, no database server.
# Run the local MCP server
npx retrace-mcp@latest
# Or add to your MCP client config:
{
"mcpServers": {
"retrace": {
"command": "npx",
"args": ["-y", "retrace-mcp@latest"]
}
}
}Your AI agent logs entries automatically as you work — decisions, bugs, people, references. You never have to open an app to write anything.
Ask your agent about past work. It searches your local journal and surfaces the right context, months later.
Retrace works best when your AI agent is instructed to proactively log context. Add these instructions to your project's .github/instructions/ or agent configuration.
Add this to .github/instructions/retrace.md in your project. Copilot will follow these instructions in every session.
# Retrace Work Memory You have access to Retrace via MCP. Use it proactively: ## When to log (call log_entry) - When a meaningful task is completed (not just file edits) - When a decision is made and WHY (e.g. "chose X over Y because...") - When a bug is found — include root cause and resolution - When context-switching to a different task - When a colleague is mentioned or involved in the work ## What to include - **content**: Write for future recall 3 months from now. Include technical specifics. - **people**: Anyone mentioned, involved, or who should know - **refs**: Ticket IDs, PR numbers, doc links mentioned in conversation - **component**: The specific service/system/module being worked on - **project**: The project name (use consistent names) ## When to search (call recall or query_journal) - When the user asks about past work or decisions - When you notice the current task is similar to something done before - Before making architectural decisions — check if there's prior context
Add to your .cursorrules file. Cursor will include these instructions in its system prompt.
# Work Memory (Retrace MCP) Always use Retrace to maintain work context: 1. LOG after completing meaningful work: - Include WHO was involved (people field) - Include WHY decisions were made - Include WHAT ticket/PR (refs field) - Include WHICH system (component field) 2. SEARCH before starting related work: - Use "recall" for fuzzy questions about past work - Use "who_knows_about" when deciding who should handle something - Use "query_journal" for filtered searches (by date, project, person) 3. Write entries for your future self — be specific about: - Root causes of bugs - Architectural decisions and rationale - People's areas of expertise - External reference IDs
Add to your agent's system prompt or MCP configuration instructions.
You have access to Retrace, a work memory system. Use it to: LOG ENTRIES when: - A task is completed (log_entry with full context) - A decision is made (include the reasoning) - A bug is resolved (include root cause) - Someone is mentioned (add them to people field) QUERY MEMORY when: - User asks about past work → use "recall" tool - Need to find expertise → use "who_knows_about" tool - Need filtered search → use "query_journal" tool Always include: people involved, ticket refs, component name, project name. Write content as if explaining to someone 3 months from now.
Entries with people, refs, and component fields populated are dramatically more useful for future queries. An entry saying "Fixed bug" is almost useless 2 months later. An entry saying "Fixed Redis TTL misconfiguration (24h→5min) causing stale responses in Forbes iOS cache-layer. Worked with Marcus, resolved MOB-1634" is instantly findable.
Local SQLite is great for one machine. When you outgrow it, upgrade to the cloud — your data syncs everywhere, and you get a full dashboard.
{
"mcpServers": {
"retrace": {
"url": "https://retrace-zeta.vercel.app/mcp"
}
}
}OAuth handles authentication automatically — no tokens to copy. Migration guide →
The local MCP server is MIT-licensed and free forever. Cloud features are paid.
$0/forever
MIT open source
$8/mo
Cloud sync + dashboard
$20/mo
Per seat
No account. No config. Just run and start coding. After a few weeks, your agents know your projects, your team, and your decisions.