meminqu — guided capture interviews for a memory store

Seven questions, asked seven different ways, written down exactly as answered. That is the whole product.

An interview cycle walks each memory domain in order. For each domain, the interviewer poses 2–4 questions in deliberately different registers — direct, reflective, structural, irreverent, sparse, temporal, contrastive — receives the answers, and captures them verbatim into that domain's raw.md. No synthesis, no interpretation, no analysis. The registers rotate across the cycle so no two domains get interrogated from the same angle twice in a row.

No prerequisites. You don't need to know anything about the project's history. The mechanism is the whole product.

(Lineage: front-facing conversion of the meminqu-memory-interrogation skill from Ryan's skill-suite, MIT. What changed is documented in CHANGELOG.md — the decoupling record.)

What it does

• Seven registers (registers.py) — the question stances, kept exactly as the skill defined them, each with a description and example stems. They are data, so a caller can supply their own set; the shipped seven are the complete original set.
• Cycle planning (plan_cycle) — assigns registers per domain deterministically: consecutive domains never share a register set, and a full cycle uses every register. No fixed order, no repetition rut.
• Verbatim entry format (entry.py) — one ### Inquiry capture — <domain> (registers: …) record per domain, questions paired with answers, answers embedded byte-identical. Skipped questions are recorded as skips, not silently dropped.
• Session runner (session.py) — the record-keeping side of the interview: domain walk, skip, explicit reassignment (material that belongs elsewhere is captured under the announced domain unless the operator explicitly reassigns it), and a completion report listing which domains got entries and where they landed.

Why it works

Most "tell me about your week" systems fail in one of two places: they ask every question the same way (so every answer comes from the same angle), or they quietly rewrite what you said on the way in (so the record is the interviewer's summary, not your words). This package fixes both in code:

• Register rotation — plan_cycle guarantees angular variety. A test asserts consecutive domains differ and a full cycle covers all seven registers.
• Verbatim preservation — format_entry adds structure around answers but never touches them. The load-bearing test feeds answers with contradictions, typos, and markdown through formatting and asserts byte-identical survival.
• Capture purity — answers go through memdate's capture(), which raises InterpretationError if the text looks like synthesis ("in summary…") instead of a raw record. The refusal is the feature: it keeps capture and interpretation from mixing in the same motion.

Installation

meminqu is a thin runner over memdate — it does not reimplement capture, config, or storage. Install the dependency first:

# from the staging tree
pip install -e ../memdate
pip install -e .

or for a quick offline run, put both src trees on the path:

PYTHONPATH=../memdate/src:src python3 -m unittest discover -s tests
PYTHONPATH=../memdate/src:src python3 examples/demo_interview.py

pyproject.toml declares memdate as a dependency honestly — this package cannot function without it.

Quickstart

from meminqu import InterviewSession, Response
from memdate import MemoryConfig

config = MemoryConfig(root="./my-memory")  # domains from config, not from here
session = InterviewSession(config)

for domain in session.domains:
    registers = session.start_domain(domain)   # announce; get suggested registers
    # ... the operator/model asks questions, collects answers ...
    session.submit_answers(domain, [
        Response("direct", "What is the key fact?", "The server migration finished Tuesday."),
        Response("sparse", "This domain in one sentence.", "Quiet, for once."),
    ])
    # session.skip_domain(domain)              # or skip it — nothing is written

print(session.render_report())

Where the formats meet

The original skill specified captures under a ## YYYY-MM-DD day-heading. The converted memdate writes a ## YYYY-MM-DD HH:MM TZ heading per entry instead. This package defers to the library's heading — one source of truth for the capture format — and the ### Inquiry capture — <domain> (registers: …) marker rides inside the entry body, where the register list stays attached to the answers it describes.

Evidence taxonomy

This package tags no evidence itself; its artifacts are raw.md entries. memdate's README maps those onto the shared ten-tier taxonomy (raw.md entries sit at OBSERVATION / CLAIM). No third taxonomy was invented.

Worked example

• examples/example-1-full-cycle.md — one full annotated cycle on neutral domains, register rotation shown.
• examples/demo_interview.py — scripted end-to-end run against a temp store; no model, no network.