How PM Brain works

May 20, 2026 · View on GitHub

A worked example, end to end. One real-shaped customer interview, six files touched, one audit trail.

The setup

You're the PM for a B2B SaaS compliance tool. Mid-market. Six active features. You've just finished a 45-minute interview with the Ops Lead at Acme Corp.

You drop the transcript into your working directory and run /ingest.

What the agent does, step by step

Step 1: Detect shape

The agent reads the artifact. Sees: speaker labels, customer name, length, question/answer structure. Classifies it as interview. Routes to ingestion/interviews/.

Step 2: Copy the source

source/interviews/2026-04-22-acme-ops-lead.md

Immutable. Read-only from this point on. The original stays in your Drive. This copy is the audit anchor.

Step 3: Tag observations

The agent creates the working-memory record:

ingestion/interviews/2026-04-22-acme.md

Inside, observations tagged with their epistemic type:

  • observation: direct quote, factual claim from the interviewee
  • interpretation: agent's framing of what the observation means
  • hypothesis: testable belief generated from the conversation
  • assumption: implicit claim the interview didn't validate

Example entry:

- **observation**: "We batch our SOC2 evidence collection on Fridays. Daily would
  break our workflow." - Ops Lead, 2026-04-22, transcript line 47
- **interpretation**: Mid-market ops leads appear to prefer weekly batch over
  real-time for compliance work. Third independent observation of this pattern
  (see also: 2026-03-15 Stripe interview, 2026-04-08 Notion interview).
- **hypothesis** (proposed): H2. Real-time alerts have negative value for the
  ops-lead persona below team size N.

Provenance stays explicit. The interpretation links to the observation. The hypothesis links to the interpretation.

Step 4: Propagate to durable layer (in parallel)

The same ingestion updates four durable destinations:

1. knowledge/users/insights.md: promotes a new insight:

"Ops leads in mid-market batch their compliance reviews weekly, not daily." Confidence: high (3 independent observations). Promoted 2026-04-22 from ingestion/interviews/2026-04-22-acme.md (with cross-links to 2026-03-15 and 2026-04-08).

Promotion threshold met: 3 independent observations of the same pattern, all from the same persona.

2. hypotheses/weekly-digest.md § Evidence: strengthens H2 (value-risk):

## H2. Value risk: real-time alerts may have negative value for ops leads

**Evidence for:** (was 0.4) → 0.7
- 2026-04-22 Acme Ops Lead: "Daily would break our workflow."  [ingestion/interviews/2026-04-22-acme.md](../ingestion/interviews/2026-04-22-acme.md)
- 2026-04-08 Notion Ops: "We don't want more interruptions."  [ingestion/interviews/2026-04-08-notion.md](../ingestion/interviews/2026-04-08-notion.md)
- 2026-03-15 Stripe Compliance: "Batch is how we operate."  [ingestion/interviews/2026-03-15-stripe.md](../ingestion/interviews/2026-03-15-stripe.md)

**Decision trigger:** if confidence > 0.8 across 5+ independent observations,
draft decision to defer real-time alerts.

Each evidence row carries a provenance tag from the canonical enum. Most here are path-typed ([ingestion/...]). The link walks to the synthesis record, which itself links to the raw source/ transcript. The vocabulary also allows non-path tags like (stakeholder-verbal, <name>, <date>), (intuition, PM, <date>), (industry-knowledge), and (chat, no artifact) for claims that legitimately exist without an artifact behind them. See hypotheses/_SCHEMA.md for the full enum.

Confidence raised. Decision trigger documented but not fired yet.

3. stakeholders/acme-ops.md § Touchpoints: logs the conversation:

## Touchpoints

- 2026-04-22: 45-min customer interview. New concern: SOC2 evidence cadence.
  Last touched: 2026-04-22. Next nudge: 2026-06-22.
  [ingestion](../ingestion/interviews/2026-04-22-acme.md)

The "Last touched" date drives the stakeholder cadence check in the weekly sweep.

4. decisions/2026-04-25-defer-realtime-alerts.md: drafted as pending:

# Defer real-time alerts

**Status:** pending (drafted by agent, awaiting PM confirmation)
**Date:** 2026-04-25
**Driver:** H2 confidence raised to 0.7. Pattern recurring across 3 mid-market
ops interviews.

**What we're deciding:** ship weekly batch alerts only for v1. Defer real-time
to v2 pending evidence from larger team sizes.

**What would reverse this:** if 2+ enterprise customers (team size > 50)
explicitly request real-time and demonstrate intent-to-buy contingent on it.

**Evidence trail:**
- [hypothesis H2](../hypotheses/weekly-digest.md#h2)
- [insight: weekly batch pattern](../knowledge/users/insights.md)
- 3 supporting interviews (linked from H2)

The agent does not mark the decision as decided. It surfaces the draft for the PM's confirmation.

Step 5: Close the loop

The agent reports back in 2-4 bullets:

Resolved: 1 source copied, 1 ingestion record created, 1 insight promoted, 1 hypothesis strengthened (0.4 → 0.7), 1 stakeholder touchpoint logged.

Drafted for your confirmation: decisions/2026-04-25-defer-realtime-alerts.md. Confidence on H2 crossed 0.7, decision trigger conditions are 80% met. Want me to mark this decided?

Surfaced: H4 (feasibility risk on the alert pipeline) still has no evidence in 30+ days. Worth a feasibility check before deferring.

Next time you /prep acme-ops, I'll surface the SOC2 cadence concern.

No dangling ambiguity. Every task ends with explicit closure.

What this gets you

Three months later, your CTO asks: "Why did we kill real-time alerts?"

You open decisions/2026-04-25-defer-realtime-alerts.md. Every evidence row carries a provenance tag. Most are path-typed [ingestion/...] links to interview synthesis records, which themselves link to the raw source/ transcripts. A few rows are tagged (stakeholder-verbal, ...) or (intuition, PM, ...). They're legitimate inputs, just ones that never had an artifact behind them, and they wear that fact. The full reasoning chain on the path-typed rows is two clicks deep.

That's the audit trail. The vocabulary is what keeps it honest when real PM work is messy.

The six verbs, briefly

/ingest is the workhorse. The other five operate on the durable layer:

  • /prep <stakeholder>: read-only at call time. Surfaces open asks, last unresolved concern, suggested questions. After the meeting, /ingest the notes.
  • /review: weekly maintenance sweep. Six checks. Edits where confidence is high; drafts where not.
  • /ideate <problem>: loads strategy + insights + hypotheses + recent decisions, surfaces 3-7 solution directions tagged with evidence.
  • /risk <feature>: 5-area risk scan. Drafts hypothesis stubs for any uncovered risk area.
  • /plan <objective>: drafts a six-block plan: what we know, assumption vs evidence, who to interview, hypotheses to open, experiments, decision points.

The maintenance sweep

Memory systems fail at month three because nothing sweeps. The weekly sweep is the most important operation in the system.

Six checks, dated report in maintenance/log/:

  1. Stale knowledge: files not updated in 6+ weeks. Still true? Archive?
  2. Stale evidence: market intel past 30-60 days, interviews past 90, stakeholder assumptions past 30, strategy assumptions past quarterly. Flags. Doesn't auto-decay.
  3. Hypothesis hygiene: actives with no evidence in 30+ days; promoteds without decisions; decisions whose reversal condition triggered; pending decisions older than 14 days with blocker impact.
  4. Stakeholder cadence + strategy tensions: high-influence stakeholders not touched in 3+ weeks; recent decisions diverging from strategy.
  5. Knowledge synthesis (compression): the highest-leverage step. Identifies recurring patterns AND recurring contradictions. Preserves minority signals. Compression is additive, never destructive.
  6. Archival sweep: shipped features inactive 90+ days; resolved hypotheses; closed asks. Before archiving anything, extracts durable lessons.

20 minutes Friday afternoon. Skip it for a month and the system rots.

What the system is not good at

  • Not fully autonomous PM. Supports judgment. Doesn't replace it.
  • Not enterprise knowledge management. One operator, one product.
  • Not maximum information capture. Deliberately throws things out.
  • Stakeholder files feel awkward at first. Writing down what your manager cares about can feel like treating people as objects. The reframe: this isn't about them, it's about your continuity. You forget. The file remembers.
  • The ad-hoc inbox tempts laziness. Every ad-hoc item gets resolved in the same session. The folder is a sorting bench, not a backlog.

See docs/architecture.md for the why behind the structural choices, and docs/testing.md for how we verify these dynamics with synthetic scenarios.