# Continual Learning

> Auto-mine session for durable memory: user preferences and workspace facts. Trigger: 'learn from this', 'update memory', 'save what we learned', or at session end.

- Skill: `aryaminus/continual-learning` (Agent Skill)
- Install (CLI): `npx skillmds@latest add aryaminus/continual-learning`
- Raw SKILL.md: https://api.skillmd.com/api/skills/aryaminus/continual-learning/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- License: Apache-2.0
- Author: aryaminus (https://skillmd.com/u/aryaminus)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/aryaminus/continual-learning

---


# Continual Learning

Auto-mine session activity for durable memory updates. CK's `ck_memory_record` is manual — this skill automates extraction.

## Do NOT use when
- During active coding (distracts from the task)
- When no meaningful session activity has occurred
- To record secrets, tokens, or transient debugging state

## What gets learned (only two categories)

1. **User Preferences**: Recurring corrections, preferred patterns, naming, tools, workflows
2. **Workspace Facts**: Stable codebase facts, architecture decisions, module ownership, constraints

## What does NOT get learned

Secrets, one-off instructions, transient state, process meta-guidance

## Workflow

1. Search `ck_memory_search` for existing memories to avoid duplication.
2. Scan session recent events (reviews, findings, decisions) for:
   - User corrections recurring more than once
   - Architecture decisions approved and not reversed
   - Workspace facts discovered and not contradicted
   - Preferences stated explicitly or demonstrated consistently
3. For each candidate: Is it durable? Reusable? Specific? If no to any, skip.
4. Record via `ck_memory_record`:
   - Preferences → type `decision`, tags `["user-preference"]`
   - Facts → type `brief`, tags `["workspace-fact"]`
5. Archive superseded items via `ck_memory_archive`.
6. If no high-signal items found: "No high-signal memory updates."

## Rules

- Plain bullet points only — no metadata blocks, confidence scores
- Cap at 12 items per section
- Deduplicate semantically — "prefer Req" and "always use Req" are the same item
- Update in place, do not create duplicates alongside old items
- Never record secrets

## Output

- `ck_memory_record` entries for new preferences/facts
- `ck_memory_archive` for superseded items
- Summary: "Recorded N preferences, M facts, archived K items."