elizaos--eliza/packages/benchmarks/memperf/AGENTS.md

memperf — Agent Guide

Desktop/server memory-benchmark harness (issue #8809). Records, per available Eliza-1 tier × modality (text, embedding, transcription, tts, vad, vision): load ms, resident RSS delta, peak RSS, tok/s or RTF, and the MemoryArbiter eviction count under a scripted co-residency sequence. Emits a JSON report on the shared METRIC_SCHEMA (shared with #8800), checks budgets.json, and exits non-zero on regression. Not registered in the suite orchestrator — run directly with node / bun.

Run

# Full harness + consolidated dashboard (results/summary/latest.md + .json)
bun run bench:memperf
node packages/benchmarks/memperf/run-all.mjs            # equivalent
bun run bench:memperf:json                              # JSON to stdout

# The measuring harness directly (TS — imports real plugin services):
bun --conditions=eliza-source packages/benchmarks/memperf/memperf-kpi.ts
bun --conditions=eliza-source packages/benchmarks/memperf/memperf-kpi.ts --json

# Limit tiers / generation length:
MEMPERF_TIERS=eliza-1-2b,eliza-1-4b bun --conditions=eliza-source \
  packages/benchmarks/memperf/memperf-kpi.ts
MEMPERF_MAX_TOKENS=64 bun run bench:memperf

--conditions=eliza-source is required — the harness imports @elizaos/plugin-local-inference source (MemoryArbiter, engine, hardware probe) under the eliza-source export condition.

Smoke test (no models, CI-safe)

# No model bundle installed → all (tier × modality) rows skip with a concrete
# reason, the co-residency self-check exercises the real arbiter fit/pressure
# eviction telemetry, and the harness exits 2 (skipped). Runs anywhere.
bun run bench:memperf

Test the harness

bun test --conditions=eliza-source packages/benchmarks/memperf/metric-schema.test.ts
bun test --conditions=eliza-source packages/benchmarks/memperf/co-residency.test.ts

# Typecheck (memperf is not a workspace package; use its standalone config):
node_modules/.bin/tsgo --noEmit -p packages/benchmarks/memperf/tsconfig.check.json

• metric-schema.test.ts pins the field set shared with #8800 (a rename/drop fails the test and must bump METRIC_SCHEMA_VERSION).
• co-residency.test.ts drives the real MemoryArbiter with synthetic sized loaders and asserts the LRU fit-path and the critical-pressure path both emit the eviction telemetry the harness counts — no models, no FFI.

Layout

Path	Role
memperf-kpi.ts	Measuring harness (real arbiter + engine + hardware probe)
run-all.mjs	Orchestrator: spawns the harness, writes the dashboard, propagates exit code
metric-schema.mjs	METRIC_SCHEMA shared with #8800 + the skipped-row builder
lib.mjs	RSS sampling, result recording, git context, budget loader
budgets.json	Per-tier peak-RSS + co-residency eviction budgets
metric-schema.test.ts / co-residency.test.ts	Schema + arbiter-telemetry tests
tsconfig.check.json	Standalone typecheck config (memperf is not a workspace package)
results/	Timestamped JSON results (gitignored; only .gitignore committed)

Notes / gotchas

• Honesty contract. A row is measured: true only on a real load+run; otherwise measured: false with a skipReason. Numeric fields are null (never 0) when unmeasured. The co-residency self-check (mode: "self-check") is never a tier metric and can never satisfy the real eviction-regression gate, but it DOES fail loudly if the arbiter stops counting evictions.
• A real measured row requires a curated Eliza-1 tier bundle (id eliza-1-*) installed via the local-inference registry — the fused libelizainference backend resolves models from the bundle layout (.../text/*.gguf), not from bare external GGUF blobs. External LM-Studio / Ollama / HF scans are deliberately NOT measured as Eliza-1 tiers.
• Exit codes: 0 pass, 1 budget/telemetry regression, 2 nothing measurable (self-check passed) — usable directly as a CI gate.
• Results write to results/memperf/latest.json and results/summary/latest.md (the results/ tree is gitignored).
• The metric schema mirrors plugins/plugin-local-inference/docs/memory-and-e2e-latency-review.md §5 and the iOS grind (plugin-capacitor-bridge/src/ios/model-grind.ts); #8809 owns the desktop/server harness + arbiter telemetry, #8800 owns the mobile surface and consumes the same METRIC_SCHEMA.
• Full overview + env reference: README.md.

⛔ NON-NEGOTIABLE — evidence, trajectories & real end-to-end tests

The binding, repo-wide standard is AGENTS.md. Read it. Nothing in this package is done until it is proven done — a reviewer must confirm it works without reading the code, from the artifacts you attach. This applies to every feature, fix, refactor, and chore here. "Tests pass" is not proof; "CI is green" is not proof.

• Record AND read model trajectories. Capture the actual inputs and outputs of the model from a live LLM — not the deterministic proxy, not a mock: the prompt, the providers/context, the raw model output, every tool/action call, and the result. Then open the trajectory and review it by hand. A captured-but-unread trajectory is not evidence (packages/scenario-runner/bin/eliza-scenarios run <scenario> --report <out>).
• Real, full-featured E2E — no larp. Every feature ships detailed end-to-end tests that drive the real path end to end. Not the happy "front door" only: cover error paths, edge/empty/invalid input, concurrency, roles/permissions, and adversarial input. A test that asserts against a mock/stub/fixture standing in for the thing under test does not count. If the real model/device/chain/connector/account is hard to reach, make it reachable — that is the work, not an excuse to mock. If the existing tests here are shallow or mocked, fixing them is part of your change.
• Screenshots + logs at every phase, plus a complete walkthrough video/run-through of the entire feature or view, start to finish (bun run test:e2e:record).
• Manually review every artifact the change touches — never just the green check: client logs (console + network), server logs ([ClassName] …), the model trajectories in and out, before/after full-page screenshots, and the domain artifacts listed below for this package.
• No residuals. No shortcuts. The goal is not "done" — it is everything done. Clear every blocker by the hard path: build the real architecture, stand up the real model/device/service, actually test it. Never leave a TODO, a stub, a stepping-stone, or a "follow-up." When unsure, research thoroughly, weigh the options, and ship the best, highest-effort, production-ready version. Keep going until every possibility is exhausted.

Artifacts → attached inline in the PR (MP4 video, JPG screenshots, logs in <details>); attach each evidence type or explicitly mark it N/A with a reason — never leave it blank. If develop moved and changed behavior, re-capture evidence; stale proof is worse than none.

Capture & manually review for this package — benchmark / eval suite:

• A real-model run (not the mock/smoke fixture) producing the score-report JSON, with the numbers inspected and the provider/model recorded.
• The per-item trajectories the harness captured, spot-reviewed for correctness — a green harness run over mock fixtures is not a result.
• The provider matrix actually exercised, and the scoring math validated against a known case.
• Failure / timeout / partial-output handling in the harness itself.