Structured Log

Write structured entries to memory/YYYY-MM-DD.md instead of free-form text. This improves recall via memory_search.

When to Use

After any meaningful action:

• Made a decision or received a decision from the founder
• Discovered a new fact (via Exa, PostHog, web, conversation)
• Found an insight or pattern
• Completed a task with a result
• Updated a hypothesis

Entry Format

### HH:MM — [TYPE] Title

What happened (1-2 sentences).

**Facts:**
- concrete fact 1
- concrete fact 2

**Concepts:** keyword1, keyword2, keyword3
**Files:** path/to/file (if changed)

Types (TYPE)

Type	When
DECISION	Decision made (by whom, why, what changes)
FACT	New confirmed fact (source required)
INSIGHT	Pattern, connection, conclusion from data
TOOL	Finding from tool usage (PostHog, Exa, web)
TASK	Completed task with result
HYPOTHESIS	Created/updated/killed a hypothesis

Rules

1. Facts must be verifiable. "Signups = 0 for 7 days" (ok). "Seems low" (no).
2. Concepts are for search. Write keywords you'd use to find this entry later via memory_search. Include: names, products, metrics, topics.
3. Files — only if changed. List only files that were created or modified.
4. One entry = one topic. Don't mix a decision + tool finding in one entry.
5. Narrative < 50 words. Context goes in facts, not prose.

Promotion to MEMORY.md

A fact deserves promotion to MEMORY.md if:

• It changes understanding of wedge, ICP, constraint
• It's confirmed 2+ times from different sources
• It affects current strategy or hypothesis

On promotion — add to the appropriate MEMORY.md section and remove outdated info.

Example

### 14:30 — [TOOL] Analytics: 41% onboarding drop-off

Analytics query revealed a high drop-off rate during onboarding.

**Facts:**
- 41% of users who start onboarding don't complete the key activation step (30d)
- Main cause: timeout on large data imports
- Affects the signup -> activated funnel stage

**Concepts:** onboarding, activation-blocker, funnel-drop, analytics, TTFV
**Files:** —