Certification Report — Agent Memory Manager

The certification state, Evidence Grade, and Trust Score below are assigned by
the Forge Proof Layer (tools/forge-proof.mjs) under the IRS_AUDITOR
standard. They are reproduced here verbatim from
proof/PROOF_DECISION.json and may not be edited by hand.

What CERTIFIED means here

All Proof Layer gates passed: required artifacts present (14/14), verification completed, evidence complete, reproduction instructions exist, zero unsupported claims, required seams disclosed, and the hostile-auditor verdict is PASS. The gate permits deliver, publish, and certify (node tools/forge-gate.mjs --action <…> exits 0).

What CERTIFIED does NOT mean

The Evidence Grade is B, not A/A+: the benchmark is synthetic (seeded corpus with a known answer key), there is no official external benchmark, and no independent reproduction or external validation has been recorded. The outcome is not claimed to be PRODUCTION_VALIDATED. The benchmarkClaimAllowed gate is false.

Disclosed seams (must be shown wherever this is presented)

1. Synthetic benchmark input (seeded corpus, known answer key) — not real traffic.
2. Default embeddings are lexical hashing, not learned-semantic.
3. Default summarizer is extractive; the Claude/LLM summarizer is a network seam.
4. Postgres + pgvector runs in-process (PGlite); the external-server path via

pg is wire-compatible but not exercised here.

1. Fleet sync runs over an in-process bus; a distributed broker is not exercised.

Full detail: proof/LIMITATIONS.md, proof/AUDITOR_OBJECTIONS.md, proof/AUDITOR_CHALLENGE.md, proof/EVIDENCE_GRADE.md, proof/TRUST_SCORE.json.

How to re-derive this report

node tools/forge-proof.mjs --outcome delivery-package/agent-memory-manager \
  --verify "node verify.mjs" --report verification-report.json
node tools/forge-proof-verify.mjs --outcome delivery-package/agent-memory-manager