Q
Why not just use RSS?
Because mmap'd model files inflate it and compressed pages vanish from it. On macOS the pair is required; a single RSS number is not a claim, it is a guess.
Glossary · Footprint vs RSS
Footprint vs RSS.
Two memory numbers, one truth on macOS.
Direct answer
On macOS, ps RSS includes shared mmap'd pages and hides compressed pages, so it can overstate or understate real memory. The kernel's phys_footprint is the honest single-number instrument. Serious memory claims therefore print the pair (footprint and RSS) with the measurement protocol, binary and date; Linux equivalents are RSS plus PSS plus Private_Dirty from /proc/<pid>/smaps_rollup. Never compare numbers across operating systems or instruments, and never quote a baseline without its peak.
In Anvaya
How We Implement It.
01Footprint vs RSSAnvaya publishes footprint/RSS pairs: CLI TUI idle 4.2 MB footprint (12.4 MB RSS), headless 2.8 MB, first frame 12 ms; Mind cold 3→6 MB, 31 MB after first queries, warm 47–51 MB, with a transient ingest burst. The Phase 1 benchmark samples the whole process tree rather than one process.
Questions
Asked About Footprint vs RSS.
Q
Why do you report the ingest burst?
Because a baseline without its peak hides the worst case. The burst is a deterministic ONNX-embed transient, not a leak, and hiding it would be the dishonest move.
Q
Can I verify these numbers?
Yes, the protocol (footprint + vmmap on macOS, smaps_rollup on Linux) is documented and reproducible; the harness's own telemetry samples RSS from the process tree at 5 s intervals.
Run Agents That Fit On Your Laptop.
25.6 MB median RSS. 25 agents ran in parallel on a Core 2 Duo with 4 GB RAM. Hundreds on your machine. Zero cloud required on the Ollama path.
Requires Rust/cargo to build from source. Linux and macOS today, Windows not yet supported. Pre-1.0, public beta. Pricing TBD.