HEL-857: approved Mem0 isolation spike - zero-model gates all pass, A/B synthetic scenarios within 26/60 real-call budget
- Fixed mem0 source commit 2c81af3 (2.2.1) in isolated venv, qdrant embedded + fastembed local zh embeddings (95.3MB, within 500MB cap)
- Gate0 transport-level persistent debit ledger: 61st call blocked pre-send, failures count, cross-process persistence, retries disabled+counted (fake transport, zero network)
- Gate1 qdrant embedded checks: keyword_search functional (Chinese segmentation gap recorded honestly), restart subprocess, clean-copy restore, no telemetry collections
- Gate2 privacy with stub LLM: P1-P5 zero leak, first-share timing F1-F6 (failed/cancel/unconfirmed no migration, ack+judged migrates told content with anchors, duplicate receipt idempotent), full backup incl. unshared internal memories restores clean
- A/B same-material same-criteria: Mem0 paraphrase recall strictly better than current lexical path (S3, model answers in Chinese), S1-S9 all pass, no regression
- Headline findings: Chinese input extracted as English memories (default upstream prompt), BM25 lexical channel inert for Chinese, ADD-only infer confirmed with our update orchestration, custom metadata needs direct payload read
- 26/60 real outbound via persistent ledger (incl. one parse-failure retry), usage/latency per row; embeddings 28/500; HEL-835 quota untouched
- Experiment branch only: no merge to main/product, no deploy, no real-profile access; handoff JSON + dual-base requirement checks pass
Co-authored-by: multica-agent <github@multica.ai>