| name | memory-bridge-operation |
| description | Manage the Hermes ↔ repo memory sync system — bridge, quality gate, compaction, health checks, and cron |
Memory Bridge Operation
Manage the Hermes ↔ repo memory sync system that propagates context across all machines.
Architecture
Memory travels with the repository via git. Hermes memory (~/.hermes/memories/) is the source of truth on ace-linux-1. The bridge script reads it, writes into .claude/memory/ in the workspace-hub repo, commits, and pushes. Every other machine (Windows licensed-win-1, new clones) gets updated context via git pull. Gateway must be running for cron jobs to fire.
Script Inventory
| Script | Purpose |
|---|
| scripts/memory/bridge-hermes-claude.sh --commit | Reads Hermes memory, injects into agents.md via BRIDGE markers, commits and pushes |
| scripts/memory/pre-bridge-quality.sh --fix | Quality gate (0-100 score) before bridge; auto-compacts if needed |
| scripts/memory/check-memory-drift.sh | Exits 1 if Hermes memory ahead of repo, 0 if in sync |
| scripts/memory/bootstrap-machine.sh | Creates ~/.claude/CLAUDE.md on new machine, OS-aware |
| scripts/upkeep/health-check.sh --save | 16-check report: gateway, cron, memory, disk, repo sync, Claude topics |
Cron Jobs
- memory-bridge-daily: 04:00 daily (job 8c797470d7d3) — runs pre-bridge-quality.sh --fix which calls bridge internally
- Gateway must be active (systemctl is-active hermes-gateway.service) or cron jobs won't fire
Memory Compaction
Hermes memory has hard limits: MEMORY.md = 2200 chars, USER.md = 1375 chars. When approaching 90%+, compact before bridging.
Manual Compaction Steps
- Check current sizes:
wc -c ~/.hermes/memories/MEMORY.md ~/.hermes/memories/USER.md
- Read current content:
cat ~/.hermes/memories/MEMORY.md
- Identify entries to remove: resolved issues > 90 days old, stale tool versions, duplicates with USER.md
- Write compacted version to /tmp/compacted-memory.md (DO NOT use write_file on ~/.hermes/ paths — overlay issue)
- Apply via terminal:
cp /tmp/compacted-memory.md ~/.hermes/memories/MEMORY.md
- Verify all critical facts preserved
- Run bridge:
bash scripts/memory/pre-bridge-quality.sh --fix
Critical Facts That Must Always Be Preserved
- /mnt/local-analysis/workspace-hub is the real mount
- ~/workspace-hub is sparse overlay — write to /tmp/ then move
- uv run on Linux, python on Windows
- digitalmodel/ is separate git repo — cd in before commits
- .legal-deny-list.yaml mandatory for doc-intelligence work
- aceengineer-strategy/ GTM context
- Hermes: 691 skills, 5 external_dirs repos
- User preferences: context parity, adversarial review, overnight batch execution
Gateway Stability (2026-04-05 findings)
The gateway DOES stay alive with zero messaging platforms — it logs "No messaging platforms enabled" then "Gateway will continue running for cron job execution" and starts the cron ticker. The real crashes are from:
- API credit exhaustion — if a cron job (e.g. gemini-overnight-batch-1) hits a 402 OpenRouter error, the gateway exits. Check
~/.hermes/logs/gateway.log for "Non-retryable client error: Error code: 402"
- Restart loop — after crash, systemd restarts it. If it again finds no platforms AND no pending cron jobs, it may exit quickly (53ms). Restart with:
systemctl restart hermes-gateway.service
- Replace flag — if gateway won't start because old PID lingers:
hermes gateway run --replace
Cron without gateway (manual fallback)
If gateway won't stay up, fire due jobs manually:
hermes cron tick
This ticks once and exits — no persistent gateway needed.
Troubleshooting
| Symptom | Fix |
|---|
| Gateway down | systemctl start hermes-gateway.service; if crashes again, check gateway.log for 402 errors |
| Bridge didn't commit (dirty submodule) | git stash && cd /path/to/repo && bash scripts/memory/bridge-hermes-claude.sh --commit && git push |
| Memory at 98%+ | Compact manually (steps above), then pre-bridge-quality.sh --fix |
| Windows missing context | bash scripts/memory/bootstrap-machine.sh && git pull |
| Drift detected | bash scripts/memory/bridge-hermes-claude.sh --commit && git push |
| Cron jobs not firing | Check gateway.log; if 402, pause offending job; if no platforms, gateway still works |
Pitfalls
- write_file/patch to ~/.hermes/ paths may hit overlay issues — use terminal to cp/mv
- Bridge git pull --rebase fails with dirty submodules — script handles via stash/restore, but heavyequipemnt-rag may still cause issues
pre-bridge-quality.sh --fix calls bridge-hermes-claude.sh --commit internally, and both can fail in dirty worktrees because the bridge runs git add .claude/memory/ and then git stash push, which stashes the staged memory changes before git commit, yielding “nothing added to commit.” Recovery: restore only memory files from the generated pre-bridge-stash, commit/push only .claude/memory/, leave unrelated untracked files untouched, then verify both drift and remote HEAD:
git stash show --name-status stash@{0} | sed -n '1,160p'
git checkout stash@{0} -- .claude/memory
git add .claude/memory
git commit -m "Sync Hermes memory bridge"
git push
bash scripts/memory/check-memory-drift.sh
git rev-parse --short HEAD && git rev-parse --short origin/main
If the stash contains non-memory paths, do not restore them during the bridge recovery unless they are explicitly part of the memory deliverable; keep the stash as backup.- Memory topic snapshots can contain literal merge-conflict markers for forensic lessons; pre-commit blocks those unless each marker line carries
CONFLICT_MARKER_FORENSIC_OK or the upstream auto-memory source is cleaned/exempted before bridging.
- Gateway crashes — check
journalctl -u hermes-gateway --no-pager -n 30 and ~/.hermes/logs/gateway.log
- Existing compact-memory.py and curate-memory.py have bugs (wrong default paths) — use manual compaction instead
- Cron output file may not exist after gateway restart
hermes cron tick works as manual fallback when gateway won't stay alive
