1Decision
The maximum number of cached processes of the Pixel targets is lowered from 1024 to 48 by setting config_customizedMaxCachedProcesses to 48 in core/res/res/values/config.xml of frameworks/base and no longer carrying the Pixel vendor overlay value of 1024 (config/device/common/overlay-inclusion.yml of vendor/adevtool), as the patch package memory/cached-process-cap in sovereignos-platform. The cached limit becomes 24 and the empty limit 24. It does not touch secure app spawning, ART, the zygote, lmkd or any hardening default; it is the second of the memory levers after ADR-150.
2Why and evidence
Owner decision 2026-10-10 (chat): accepted with MEM-011 still NOT_RUN and the behaviour above 48 cached processes untested; both are measured on stallion after a user candidate. Owner choice 2026-10-10 (chat) of 48 among 64, 48 and 32.
Measured on lynx 2026100801 (MEM-007, MEM-008): with the twelve launcher apps opened first the anon demand fell from 5451 MB (1024) to 4643 MB (48), against 5059 MB at 64 and 4248 MB at 32, and all twelve apps stayed alive at 64, 48 and 32 while 24 lost the clock and the gallery. The last used app and every foreground, visible, perceptible and persistent process were never killed (OomAdjuster.updateAndTrimProcessLSP, verified in source).
A killed clock app still fires its alarm (MEM-009) and the AirPods connection starts the process by broadcast, builds the session and shows the sheet with no app process (MEM-010). A user build cannot change the value live, so a wrong value costs an image.
Source: ActivityManagerConstants.updateMaxCachedProcesses reads the DeviceConfig flag, else config_customizedMaxCachedProcesses, and the Pixel vendor overlay of lynx and stallion sets that resource to 1024 and overrides the framework value, so the earlier note that DEFAULT_MAX_CACHED_PROCESSES is the live default was wrong.
3Consequences
A killed cached process starts cold when it is used again, and every restart after a kill brings a new private copy of about 43 MB, so a process started often by broadcasts can cost more than it frees; this is measured nowhere yet (MEM-011).
The cap acts only ten minutes after the last user unlock (any user's unlock resets the window: UserController calls setLastUserUnlockingUptime, read by OomAdjuster.shouldKillExcessiveProcesses), so the boot burst is not reduced. The inclusion list is an allow-list read by the prebuilt arsclib, whose source is not in the tree: how a later factory image with another value of this resource is treated by adevtool generate-all is not verified.
Both patches are landed on the build server (d9665a56 in vendor/adevtool, eb4138d4 in frameworks/base) and in the package.
They are built into lynx image 2026101001 together with ADR-150's 100% (MEM-012: the built framework-res has 48, the vendor overlay none); flashed on lynx and read there (MEM-013: mCustomizedMaxCachedProcesses=48, CUR_MAX_CACHED_PROCESSES=48; MEM-014: twelve launched apps stay alive; MEM-015: the AirPods connection with no process).
On 2026-10-10 the entry was removed from the generated vendor overlay files of lynx and stallion (no overlay of either device sets it any more, read back) and both adevtool pins point to d9665a56; copies of the changed files are kept outside the overlay res/ directories (Soong compiles every file in them).
The inclusion change is not exercised by adevtool generate-all (it needs tens of GB of factory images): after the build, dumpsys activity settings must show mCustomizedMaxCachedProcesses=48 and CUR_MAX_CACHED_PROCESSES=48.
4Review when
dumpsys shows 1024 after the build; LOW_MEMORY exits do not drop or process starts per hour rise sharply on stallion (MEM-011); an app the owner uses is killed often enough to be noticed; an upstream change of the resource name or of the vendor overlay filter.
Read from docs/governance/decision-log.md, entry ADR-151.