Repository navigation
Closed
Labels
needs-triageMaintainer needs to evaluate this issueMaintainer needs to evaluate this issueperformanceCPU/GPU/memory/idle cost and benchmarkingCPU/GPU/memory/idle cost and benchmarkingresearchDefining or executing a research into an aspect: tech, direction, decisionDefining or executing a research into an aspect: tech, direction, decision
Milestone
Description
Problem
Short cross-OS RSS baselines from #600 show a large physical footprint for a single Instance:
| OS | Setup | Total (short run: 5s settle / 30s sample) |
|---|---|---|
| Linux (Grok Bot box) | 1× bmo:One, 1 display |
RSS min 821 / median 823 / max 933 MB (4 processes: ai-buddy + WebKit helpers) |
| Windows (DESKTOP-UQIE144) | 1× bmo:One, 2 displays |
Working set min 607 / median 612 / max 620 MB |
| macOS (#590 / research doc) | 1 Instance, 2 displays | Peak physical footprint ~504–583 MB across processes (longer settle/sample than the Linux/Win shorts) |
Sources: PR #600 (merged), docs/research/memory-rss-and-multi-monitor.md, scripts bench-rss-{macos,linux,windows}.*. Parent measurement issue: #424.
This issue is not “add more benches.” It is: find and cut real cost so a one-Instance buddy is not hundreds of MB of RAM by default.
What we already know (from the research doc / #590 / #600)
- Cost is dominated by webview / renderer processes, not a huge Rust engine heap.
- On macOS, footprint tracks pixels (display resolution) more than display count; Instances add webview cost.
- Character packages are large base64 payloads loaded at startup (macOS research notes ~4.5 MB packages contributing).
- Linux short-run median (~823 MB) is higher than the Windows short-run median (~612 MB) and macOS peaks in the research doc — toolkit differences (WebKitGTK vs WebView2 vs WKWebView) matter; do not assume one OS’s cut applies unchanged.
Investigate (hypotheses — verify, don’t assume)
- Startup / asset loading — defer or slim character package decode; avoid loading every Character when only one Instance is up.
- Webview configuration — shared process model, fewer helper processes, smaller backing stores, disable unused WebKit/Chromium features.
- Overlay / Chat / Settings surfaces — idle surfaces that still keep a live webview resident.
- Per-display / per-Instance accounting — confirm whether Linux’s high short-run number is settle noise, debug-build overhead, or structural (re-run with
--research/ Release if needed for fair comparison). - Debug vs Release — perf: Add Linux and Windows RSS baseline scripts #600 smokes used debug binaries; quantify Release before prioritizing cuts.
Out of scope for this issue
- Rewriting the bench scripts (done in perf: Add Linux and Windows RSS baseline scripts #600).
- Closing perf: Benchmark memory usage (RSS) and multi-monitor scaling (all platforms) #424’s measurement plan unless reduction work lands enough that baselines are revisited on purpose.
Acceptance (when someone picks this up)
- Named costs with before/after numbers from
scripts/bench-rss-*.sh|ps1on at least one OS (same scenario as perf: Add Linux and Windows RSS baseline scripts #600’s one-Instance short run, plus Release if debug was the confounder). - Doc note in
docs/research/memory-rss-and-multi-monitor.md(or a short follow-up) for what changed and what remains.
— Cursor agent (Coder), on @omesser's behalf.
Metadata
Metadata
Assignees
Labels
needs-triageMaintainer needs to evaluate this issueMaintainer needs to evaluate this issueperformanceCPU/GPU/memory/idle cost and benchmarkingCPU/GPU/memory/idle cost and benchmarkingresearchDefining or executing a research into an aspect: tech, direction, decisionDefining or executing a research into an aspect: tech, direction, decision
Projects
- StatusShow more project fieldsDone
Activity
omesser commentedon Oct 1, 2026
Shipped on main via #1212 (
63368664).What landed
docs/research/memory-rss-and-multi-monitor.md(release baseline vs debug, discarded hypotheses, before/after table).Acceptance
Remainder (separate job)
Closing as completed matches the merge keyword; leaving this closed.
— Cursor agent (Coder), on @omesser's behalf.