feat(benchmarks): rank real memory adapters

[drewstone]

What changed

• Adds runMemoryAdapterBenchmark() to run memory cases through actual AgentMemoryAdapter implementations and persist a ranking report.
• Adds a 12-case first-party memory lifecycle benchmark covering ingest, recall, temporal updates, stale forgetting, reasoning, summarization, recommendations, and multi-party attribution.
• Adds adapter-backed local contenders: createNoopMemoryBenchmarkAdapter() and createInMemoryBenchmarkAdapter().
• Documents how Neo4j Agent Memory, Mem0, Zep, Letta, Graphiti, vector stores, or product-owned memory systems enter through the same adapter interface.

Real run

• Command: pnpm exec tsx -e ... runMemoryAdapterBenchmark(...)
• Artifact: /tmp/agent-knowledge-real-memory-ranking-final/memory-adapter-ranking.json
• Result: in-memory ranked feat(readiness): bridge knowledge index to agent-eval #1 with mean score 0.9375 and pass rate 0.9167 over 12/12 cells; no-memory ranked feat: add source freshness readiness audit #2 with mean score 0.25 and pass rate 0 over 12/12 cells.
• Confirmed raw miss: /tmp/agent-knowledge-real-memory-ranking-final/in-memory/first-party_memory-lifecycle_allergy_0/cached-result.json shows the in-memory adapter retrieved stale “no food allergies” memory, scoring 0.25 on that case.

Verification

• pnpm install --frozen-lockfile --prefer-offline
• pnpm exec tsc --noEmit
• pnpm exec vitest run tests/benchmarks.test.ts — 8 passed
• pnpm lint
• git diff --check
• pnpm exec vitest run — 267 passed, 12 skipped
• pnpm build

[tangletools]

✅ Auto-approved drewstone PR — a008ddcd

This PR was opened by the trusted drewstone account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: drewstone_author · 2026-07-08T18:05:31Z}