Mortgage QA Memory MCP

Repository: C:\Repo\mcp-memory
Status: Design & templates + working POC/MVP implementation. See IMPLEMENTATION.md for the runnable packages/* monorepo, docs/INTEGRATION.md, and docs/DEFINITION-OF-DONE.md.

A from-scratch design for Playwright QA automation with a custom Memory MCP, tiered retention, and mortgage compliance audit — adapted from DoorDash's agentic memory architecture and Salesforce Agentic Memory patterns.

See PROJECT-CONTEXT.md for how this repo was assembled and AGENTS.md for agent working rules.

Related MCP server: Mochi

Who this is for

• Platform / QA engineers building internal AI tooling on Cursor, Gemini gateway, KB MCP, and Azure MCP

• Mortgage technology teams that need QA intelligence without creating a second store of loan data or NPI

• Teams evaluating Playwright MCP + a QA memory expander they control end-to-end

Document index

Reference artifacts

Executive summary

The problem

QA teams using AI agents (Cursor + Playwright MCP) face two opposing forces:

1. Agents need memory — flakiness history, journey maps, locators, environment quirks — or they rediscover the same failures every session.

2. Mortgage teams must not hoard sensitive data — NPI, raw snapshots, prompts with borrower fields, and unbounded long-term storage create compliance and security risk.

DoorDash's production memory architecture (see diagram in doc 02) solves a similar problem for consumer personalization by inserting policy enforcement before save and separating generation → pipeline → storage → retrieval.

This guide adapts that pattern for internal mortgage QA automation.

The solution: Mortgage QA Memory (MQM)

What we explicitly do not build

• Full loan file intelligence (buy Ocrolus / vendor doc AI)

• Long-term storage of a11y snapshots, prompts, or network bodies

• Agent-driven Playwright in production CI (deterministic tests only in CI)

• Unapproved agent writes to curated journey/locator registries

Recommended build sequence

Week 1: Policy + Playwright reporter + SQLite (read-only MCP)
Week 2: Journey YAML + compliance checkpoints + Cursor skill
Week 3: Playwright MCP local triage + audit client
Week 4: CI artifact + purge jobs + golden eval set

See 06-build-from-scratch.md for full detail.

Architecture at a glance

flowchart TB
  subgraph gen [Memory Generation]
    CONV[Session notes Tier 0]
    CI[Playwright CI reporter]
    EVAL[Eval platform]
  end

  subgraph pipe [Shared Save Pipeline]
    SAN[Sanitize]
    EXT[Extract facts]
    DED[Dedupe and merge]
  end

  subgraph pol [Memory Policy - pre-save]
    RET[Retention rules]
    PII[PII deny patterns]
    PERM[Agent write permissions]
  end

  subgraph store [Storage]
    T0[(Session Redis 8h)]
    T1[(Operational SQLite 30d)]
    T2[(Curated journeys git)]
    AUD[(Audit metadata 365d)]
  end

  subgraph tools [MCP Tooling]
    MQM[mortgage-qa-memory MCP]
    PW[Playwright MCP]
  end

  subgraph agents [Agents]
    CUR[Cursor QA agent]
    PRA[PR assistant]
  end

  CONV --> SAN
  CI --> SAN
  SAN --> EXT --> DED --> pol
  pol -->|allow| T1
  pol -->|allow| T0
  pol -->|human approve| T2
  pol -->|deny| BLOCK[Blocked + audit]
  MQM --> T0 & T1 & T2 & AUD
  CUR --> MQM & PW
  PW --> AUD
  EVAL <--> T1
  PRA --> MQM

Key design decisions (locked)

Related external references

• DoorDash Ask DoorDash / InfoQ summary — agentic memory + MCP + eval at scale

• Playwright MCP docs — browser automation via MCP

• flakiness-knowledge-graph-mcp — reporter + SQLite + MCP pattern to fork

• Fannie Mae LL-2026-04 — AI governance for seller/servicers (effective Aug 6, 2026)

• Blend Autopilot MCP — lending MCP reference architecture

Next steps

1. Review 05-data-retention-and-privacy.md with security / compliance

2. Customize policies/mqm-policy.yaml for your staging URLs and retention windows

3. Follow 06-build-from-scratch.md Week 1 checklist

4. Add three journey YAML files for your highest-traffic borrower flows