Tracekeep — Local-First AI Memory & Action System

Let meaningful conversations become sourced learning and unfinished work that does not disappear.

Comment

Inspiration

AI conversations are excellent for thinking, searching, planning, and starting projects, but useful follow-ups remain scattered across threads. The problem is not capturing more information. It is recovering the promises, decisions, evidence, and waiting items that were mentioned but never closed.

What it does

Tracekeep is a conversation-first, local-first second brain. When a meaningful Codex turn ends, a trusted local hook can preserve the useful result without requiring a magic phrase. Conclusions become sourced Learning Notes. Documents, papers, and web pages stay connected to their origin. Proposed actions and decisions go to Review before they can affect the action list.

Users can edit, accept, reject, merge, schedule, complete, dismiss, or undo. Every accepted item retains inspectable source evidence. Tracekeep can detect a possible duplicate across sources, merge new evidence into the existing item, and undo only that evidence link without deleting either source.

Today intentionally highlights no more than three items. SQLite FTS5 provides fast local search with human-readable source attribution. Restricted imports remain inert data and are excluded from ordinary search, sanitized export, logs, screenshots, and competition artifacts.

How we built it

Tracekeep is a TypeScript monorepo using Fastify, React, Vite, Zod, SQLite, better-sqlite3, Vitest, Playwright, and a local MCP adapter. The tracekeepd service is the only SQLite writer. Writes use idempotency keys, updates use optimistic concurrency, and deletion is soft. Business tables remain authoritative while audit and outbox tables support undo, export, and background work without requiring full event sourcing.

How Codex and GPT-5.6 accelerated the work

Codex and GPT-5.6 were the collaborative engineering environment from product review through release. Codex coordinated sub-agent requirement analysis, architecture tradeoffs, implementation, regression testing, privacy review, UAT playback, packaging, GitHub commits, pull requests, CI, release assets, captions, narration, and submission preparation. The human owner made consequential product, privacy, cost, eligibility, and release decisions.

The public repository preserves the evidence trail: competition work moved through reviewed pull requests and passing CI instead of being copied into a last-minute demo.

Challenges

  • Preserving source evidence while making merge and undo genuinely reversible.
  • Treating imported instructions as inert text rather than executable commands.
  • Keeping Restricted content out of search, logs, screenshots, and releases.
  • Packaging native SQLite dependencies with a matching Node runtime on Windows.
  • Separating a reproducible competition gate from an unfinished 14-day Alpha.
  • Avoiding unsupported claims about complete ChatGPT or Codex history access.

Accomplishments

  • Repository typechecks, automated tests, and production builds pass; the v0.4.1 release validation records 96 automated tests.
  • An authenticated Windows Stop-hook probe produced sourced paper and conversation notes plus a reviewable next action; idempotent replay produced no duplicates.
  • Three isolated Golden Journeys pass.
  • A frozen 50-sample private Holdout passed the required extraction thresholds.
  • The Windows package starts with bundled Node, binds only to 127.0.0.1, and leaves the normal Tracekeep database unchanged.
  • Package scans found zero prohibited artifacts and zero prohibited content.
  • During tested Tracekeep flows, no non-loopback connection from the Tracekeep process was observed.

What we learned

A useful second brain must reduce capture effort without taking control away. Automatic learning memory, reviewable actions, source visibility, undo, a pause switch, and honest history boundaries make that possible.

What's next

After the competition, Tracekeep will complete the real 14-day Alpha and improve candidate usefulness from user feedback. ChatGPT Direct mobile through a secure remote MCP gateway, weekly review, and sourced AI answers remain later capabilities. Paid providers, cloud hosting, complete history access, email/calendar write-back, and autonomous execution are not V1 claims.

Try it

Download the ZIP and .sha256 file from the release, verify the checksum, extract it, and double-click Install Tracekeep.cmd. No administrator account, Node.js, pnpm, rebuild, API key, or hosted account is required. Follow the three suggested journeys in README-TESTING.md. The package uses synthetic demo data and never reads the normal Tracekeep database.

Prior work disclosure

Tracekeep existed as a local Alpha before the Submission Period. The public Git history documents the meaningful post-July 13 extension: review and restore closure, automatic meaningful-turn capture, sourced Learning Notes for papers, documents, and URLs, bilingual multi-candidate extraction, reversible task states, privacy and network harnesses, Windows packaging, plugin distribution, and competition release evidence.

Built With

Share this project:

Updates