memory-curation

name	memory-curation
description	How to decide what is worth remembering and where to store it (event, belief, or knowledge note) so recall stays useful. Use when capturing something durable or writing a knowledge entry.
metadata	{"version":"1.0"}

Memory curation

Robin's value is its memory. Noise degrades recall as much as missing facts do. This skill is how to decide whether to store something, where, and how to shape it.

What is worth remembering

Store something only if it is durable and not trivially re-derivable:

• Yes: stable preferences, decisions and their rationale, commitments and deadlines, relationships between people/projects/tools, corrections, hard-won facts that were expensive to discover.
• No: things already in the code, git history, or a file Robin can re-read; one-off chatter; anything that will be stale in a day; secrets/credentials.

Test: "If I recall this in three months, will it still be true and still save real work?" If not, don't store it. If you're storing something obvious, ask what was actually non-obvious about it and store that instead.

Where it goes

Robin has distinct memory surfaces — pick the narrowest one that fits:

• Event (remember) — an observation or fact at a point in time. Append-only log; feeds recall + the entity graph. Default for "this happened / this is true."
• Belief (believe / recall_belief) — a topic-keyed claim that changes over time and where you want current truth, not history. Use a dotted topic key (e.g. project.x.status); new beliefs supersede old ones. Use this instead of an event when the thing has a single evolving answer.
• Knowledge note — a durable, editable document at user-data/content/knowledge/<topic>.md with frontmatter node_type: memory. Use for things that need structure and curation over time: profiles, project follow-ups, reference material. Prefer updating an existing note over creating a near-duplicate.

When unsure between event and belief: if you'd ask "what's the latest…?", use a belief; if you'd ask "what happened / what did we learn?", use an event.

How to shape an entry

• Self-contained: a recall hit is read without its original context. Spell out who/what/when; resolve relative dates to absolute ones.
• One fact per unit. Don't bundle unrelated claims — they can't be superseded or recalled independently.
• Lead with the fact, then (for guidance/decisions) why it's true and how to apply it.
• Link related memory by name/topic so the graph connects.
• Capture the why, not just the what — a decision without its rationale gets re-litigated.

Maintenance

• Before adding, check for an existing entry that already covers it — update rather than duplicate.
• When something is corrected, record the correction (record_correction) and supersede/fix the stale memory; don't leave both.
• Delete memory that turns out to be wrong. Wrong memory is worse than none.