Where This Idea Came From

I kept re-typing the same content rules into scheduled tasks that ran without me, while Claude's memory already had every one of those rules written down.

‍

Problem

Claude's memory held a couple dozen files on my work: how things get done, the rules I follow, which decisions were settled. It was the most complete record of my working context anywhere, and only Claude could read it. My other agents work from files in Fulcra, so the same rules got copied over by hand and drifted. I didn't want to move the memory, because Claude updates it on its own. I needed a copy that stayed current without me.

‍

Steps to Solve

• I wrote a sync prompt and ran it by hand once, and it copied every memory file into /agent/claude/memory/ in Fulcra, along with a manifest and a log.
• I turned the prompt into a scheduled task that runs once a day.
• I found that the log was missing three days in the first week, which meant the schedule had skipped those days without any error; the one line each run writes was the only way I knew.
• I saw a few days later that one write came back short, a single dropped line, and the run caught it by comparing sizes and rewrote the file.
• I tested a tighter version of the prompt, and the first run lost its connection to Fulcra partway through a write.
• I made two rules explicit after that: rewrite any file that comes back short, and only update the manifest for writes that succeeded, so a failed file gets retried. The next run picked up the failed file and copied it cleanly.

‍

Solution

The solution is one prompt, run as a daily scheduled task. It keeps a manifest of each file's timestamp and size so a run only copies what changed, writes one log line per run, and only ever copies in one direction, so nothing in Fulcra can touch the original.

Unattended sync. Don't ask questions or send messages; do the work and stop.

Mirror my Claude memory into my Fulcra account at /agent/claude/memory/. Claude's memory is the source of truth: only ever write memory → Fulcra, never the reverse.

1. List every memory file with its size and "updated" timestamp.
2. Read /agent/claude/memory/manifest.json from Fulcra ({"files": {path: {updated, size}}}). If it's missing, treat every file as new.
3. For each memory file that is new, or whose timestamp or size differs from the manifest: read it and write it to /agent/claude/memory<same path> as text/plain, exactly as read with its frontmatter, minus the "[updated: …] [version: …]" line the read tool adds. If Fulcra reports a size smaller than the memory file's, write it again in full (one byte larger is normal).
4. Never delete anything in Fulcra. If a manifest entry no longer exists in memory, just note it in the log.
5. Rewrite manifest.json (application/json): update the timestamp and size only for files that wrote successfully, so failed ones get retried next run; set sync_meta.last_full_sync to now (UTC) and leave the rest of sync_meta alone.
6. Append one entry to /agent/claude/memory/_sync-log.md: UTC time, files copied and their paths (or "no changes"), and any failure. Keep the newest 30 entries.
7. Don't retry a failed tool call. If the memory tools fail or Fulcra is unreachable, log a one-line failure note if you can and stop.


‍

Result

Every one of my Claude memory files is now in Fulcra, and the log shows every change since the first sync. Any agent connected to my account can read them with the same read_file call it uses for everything else, without access to Claude. The same pattern works for any agent that keeps its memory as files it can list and read.

The result: the agent that knows my work best is no longer the only one that can read what it knows.