fp-schema-memory

FP: Schema Memory

Remember patterns, not everything.

Schema memory replaces semantic memory. It does not store large context, repository summaries, or raw conversation history. It stores reusable work patterns discovered from evidence.

Use schema memory to answer:

• What kind of task is this?
• How does this kind of task usually fail?
• What context is usually useful?
• What context is usually waste?
• What verification pattern works?
• What stop condition prevents drift?

Use when

• An evidence ledger and adaptive improvement report reveal a reusable pattern.
• The same class of work has appeared more than once.
• A task type has predictable failure modes.
• A context budget pattern can reduce future context use.
• A verification pattern can prevent repeated mistakes.
• A workflow is becoming stable enough to standardize.

Do not use when

• There is no evidence ledger.
• The insight is a one-off preference.
• The pattern is speculative.
• The pattern would increase context without reducing risk.
• The agent wants to store raw memory instead of a reusable schema.

Goal

Produce or update a schema memory card.

A schema memory card captures a reusable pattern for a class of work:

• trigger
• problem pattern
• common failure modes
• recommended execution pattern
• context budget pattern
• verification pattern
• files or modules usually involved
• files or modules usually avoided
• evidence required
• promotion history

Procedure

1. Read the evidence ledger and adaptive improvement report if available.
2. Identify the class of work.
3. Extract the repeatable pattern, not the incidental details.
4. Identify common failure modes.
5. Identify the smallest useful context pattern.
6. Identify the verification pattern that proved the work.
7. Identify stop conditions that would prevent drift.
8. Decide whether to create, update, or reject the schema.
9. Keep the schema short enough to be reused in a future execution brief.
10. If the schema relates to other schema or lesson cards, populate the related-schemas YAML frontmatter with typed edges before finalizing. Use archive/templates/memory-graph-traversal.md to check blast-radius effects.
11. Populate the task-types YAML frontmatter with 3-5 keywords that describe the class of work. These keywords enable cluster retrieval via memory-graph.js. Use compact, grep-friendly terms: prefer ["bug", "validation", "auth"] over ["validation logic for authentication tokens"].
12. If this is a Map of Content (MOC), set is_moc: true, leave task-types empty, and use informs edges to list sub-cards. Create a MOC when 3 or more existing cards share a theme.
13. For Folgezettel sequences, use next and previous edges to capture the narrative order of your thinking. These are NOT semantic edges — they capture trajectory, not argument structure.
14. Follow the Zettelkasten conventions in archive/templates/zettelkasten-conventions.md: atomicity (one pattern per card), bidirectional links, MOC at N≥3, refinement pipeline, serendipity traversal, and card size constraints.