feat(mem): config-driven allocator trim with telemetry and lifecycle coverage

[aider4ryder]

Summary

Config-driven glibc malloc_trim for long-lived Hermes processes (gateway, TUI gateway, slash_worker).

Changes

• hermes_cli/mem_trim.py: trim_memory() with configurable cooldown, RSS snapshot telemetry, and forced-trim INFO logging
• gateway/run.py: periodic trim in gateway housekeeping loop
• tui_gateway/server.py: trim in idle reaper (~every 5 min)
• tui_gateway/slash_worker.py: trim on turn boundary
• run_agent.py: force trim on agent close
• hermes_cli/config.py: context.memory_trim config section

Config

context:
  memory_trim:
    enabled: true
    cooldown_seconds: 60.0
    log_every_n: 1
    info_log_min_delta_mb: 0.0

Unsupported platforms (non-Linux/non-glibc) are safe no-ops.

Testing

• 12 unit tests
• CSA tier-4 reviewed (4 rounds; 0 HIGH/MEDIUM/CRITICAL)
• Deployed and smoke-tested on production machine

Supersedes #63708 + #64591.

[alt-glitch]

This was generated by AI during triage.

Related: #63708 and #64591 cover allocator-trim infrastructure and the gateway reaper. This PR adds broader lifecycle hooks; please choose a consolidation path rather than treating it as a duplicate.

[tonydwb]

Code Review Summary

Verdict: Comment

Memory subsystem change touching 9 files with 644 additions. The PR adds config-driven allocator trim with telemetry and lifecycle coverage. While the description sounds reasonable, the broad scope across the memory subsystem warrants human review to confirm no unintended interactions with existing memory management logic.

---

Reviewed by Hermes Agent

[teknium1]

Thanks for consolidating the allocator-trim work and retaining the config.yaml mechanism from the related PRs.

Problems

• run_agent.py:3627 adds trim_memory(force=True), but AIAgent.close() recursively calls child.close() for every active child (run_agent.py:3590-3599). The helper's force path bypasses its cooldown, so a teardown with N children performs N+1 full collections/trims. Please retain rate limiting for nested closes or designate one outer teardown trim.
• The new context.memory_trim controls are not documented. website/docs/user-guide/configuration.md:812-830 currently documents only context.engine, while this PR changes no docs file.
• The tests do not cover the new AIAgent.close() or slash-worker lifecycle hooks.

Suggested changes

• Add lifecycle tests for those two hooks, including nested-child cooldown behavior.
• Add configuration documentation for defaults and Linux/glibc-only behavior.

Automated hermes-sweeper review.

[teknium1]

force=True bypasses the helper cooldown, while close() recursively calls child.close() for every active child. A parent teardown can therefore run a full collection/trim once per child plus once for the parent. Please preserve rate limiting for nested closes or arrange for only the outer teardown to force a trim.

[kshitijk4poor]

Thanks @aider4ryder — this is exactly the comprehensive trim design the earlier per-site attempts needed (it already superseded #63708/#64591, which were closed pointing here). The base had drifted (DEFAULT_CONFIG moved to config_defaults.py, housekeeping/close/reaper anchors all shifted), so I re-anchored it onto current main and salvaged it into #76905 with your authorship preserved via cherry-pick; all 5 lifecycle hooks and every test carried over (505/505 gateway-server tests green). Closing in favor of the salvage.

[aider4ryder]

FYI — this already landed upstream. It was salvaged into official main as PR #76905 (da43a8527b47 feat(mem): config-driven allocator trim with telemetry and lifecycle coverage), with follow-ups 3fad8fdc45b9 and c575351d9ac5, touching hermes_cli/mem_trim.py, gateway/run.py and tests/gateway/test_memory_trim_housekeeping.py. Superseded by upstream.