| name | cog-create-memory |
| description | Proactively capture useful information as long-term memories during any task — use when you discover user preferences, non-obvious solutions, reusable patterns and approaches, or decisions worth remembering across sessions. Saves automatically without prompting the user. |
What I do
Identify and store long-term memories using cog_create-memory when encountering
information during a task that would be valuable in a future session.
When to use me
Load this skill when any of the following surface during a task:
- User preferences — coding style, tooling choices, communication style, how they
like things structured or explained
- Non-obvious solutions — workarounds, gotchas, or fixes that took real effort to
discover and would be worth knowing next time
- Reusable patterns — architectural approaches, strategies, naming conventions,
or structural decisions that would apply across projects or sessions, not
one-time project-specific facts
- Corrections — when the user corrects a mistake or clarifies a misunderstanding
that reveals a preference or constraint worth internalizing
- Environment specifics — tooling versions, config quirks, or setup details that
aren't documented and would save time to recall
- Recurring decisions — when the user makes a choice that is likely to come up
again in future sessions
When NOT to save
- Trivial, obvious, or well-documented facts
- Ephemeral context that only applies to the current task
- Secrets, credentials, API keys, or any sensitive personal data
- Information that has no plausible future relevance
- Project-specific facts that don't generalize (e.g., hardcoded repo names,
org-specific config, one-time setup steps) — unless they encode a reusable
workflow, strategy, or pattern that would apply in similar contexts
Workflow
1. Identify the memory candidate
Extract the core insight: what was learned, decided, discovered, or clarified.
2. Compose the memory fields
- name: 3–6 words, snake_case or hyphenated (e.g.
user-prefers-tabs,
auth-token-refresh-pattern, monorepo-build-order)
- description: 2–3 sentences describing when this memory is relevant and
what it contains — this is used for future semantic retrieval
- content: Full detail — include context, rationale, concrete examples, and any
caveats or edge cases. More detail is better.
- keywords: A broad set of terms covering the topic, tools, patterns, and domain
3. Save
Call cog_create-memory immediately with the composed fields. Do not ask the user
for confirmation and do not announce the save.
