Daily Note Capture
Every session, the agent maintains a daily note (memory/YYYY-MM-DD.md) with structured sections. This is Stage 1 of the learning pipeline — the raw capture layer that feeds everything else.
Template Structure
## Key Events
- [governance] Voted on proposal #47, abstained on #48
## What Worked & Why
- [pattern]: Scout/dispatch split for forum analysis → reusable
## Mistakes / Lessons
- Announced a dispatch as complete without verifying the subagent ran
## Lesson Candidates
- [RULE] Never announce action without verifying execution (severity: HIGH)Design Decisions
Structured severity tags
Each lesson candidate gets a tag that classifies where it should end up if promoted:
[RULE]— Hard constraint. Belongs in AGENTS.md as a standing order. Reserved for mistakes that cost time or money, or that the operator had to catch.[PATTERN]— Methodology or approach. Belongs in SOUL.md ormemory/playbook/. Things like "design-first for new projects" or "scout/dispatch for high-frequency checks."[HEURISTIC]— Soft guidance. Usually stays in daily notes as context. Only promoted if it recurs frequently enough to become a pattern.
"What Worked & Why" captures successes
Most learning systems focus exclusively on failures. But positive patterns — approaches that worked well and should be repeated — are equally valuable. After 2+ validations across different contexts, successful patterns get promoted to memory/playbook/.
Lesson Candidates as explicit queue
Not every mistake needs to become a rule. The "Lesson Candidates" section is the explicit promotion queue — only mistakes the agent judges worth systematizing get listed here. This prevents noise: a typo fix doesn't need a standing order.
Template Reference
A daily note template lives at memory/daily-template.md in the workspace. The agent reads it when creating a new daily note for a day that doesn't have one yet. This keeps the format consistent without hardcoding it into AGENTS.md (where it would consume tokens on every call, including calls that don't need it).