Social Memory — who people are

What it does#

Everyone Bea has ever seen gets a tally. Only the people who made themselves matter get a card. That split is what lets her recognise a regular in a chat of thousands without the prompt filling up with strangers.

src/core/skills/social/
├── social.py   SocialMemory — the per-batch hook and two tools
└── people.py   pure decisions: should_promote, promotion_reason

The storage is in bea.db: identities, roster, roster_sessions, people, facts.

The two tiers#

A tally (RosterEntry) is a row of counters: how many messages, how much donated, whether there was ever a 1:1, how many distinct sessions. It carries no generated content, so it costs one INSERT per message and is affordable for every chatter that ever types.

A card (PersonCard) is real memory: names across platforms, facts, and how she feels about them. It is generated, so it is only worth minting for someone who earned it.

def should_promote(entry) -> bool:
    return (
        entry.donation_total > 0                              # money
        or entry.marked_by_bea                                # she decided to
        or entry.had_1on1                                     # a real 1:1
        or entry.session_count >= REGULAR_SESSION_THRESHOLD   # a regular (3+)
    )

session_count is COUNT(*) over roster_sessions, a join table rather than a counter — "three distinct sessions" is what makes a regular, and a counter that can double-count would mint cards for one-time visitors.

The per-batch hook#

context_for(batch) does three things at once, on every batch:

for each author in the batch (skipping the owner):
    ├─ meta["tallied"]?  ──► already counted by a high-volume surface, skip
    ├─ roster.record(...)     the sighting
    ├─ _maybe_promote(entry)  mints a card the moment the rule fires
    └─ collect their card

→ "[WHO YOU'RE TALKING TO]\n- **marco** (you: he's alright): builds redstone…"

Capped at MAX_CARDS_INJECTED people so a crowded room cannot take over the prompt. Cards are injected rather than looked up, so she recognises whoever is in front of her without spending a tool call to ask.

The tallied flag matters: Twitch counts every message itself, including the ones the attention gate filtered out. Counting again here would double them.

Tools#

remember_person is also armed on written channels — it is the one memory action that makes sense while she is texting.

Where cards come from#

Every source mints through one function (promote_entry in people.py): an identity whose display name exactly matches a card links to that card instead of minting a second one, when either

• the identity shared a session with the card, or
• the card is known only by name (is_placeholder): every identity on it is a synthesized named:<name>, or it has none left. It has no sessions to share, so the first account wearing exactly its name is taken to be that person. Once it has an account behind it, it is no longer a placeholder.

Near matches, and two accounts that never shared a session, stay separate — two different humans with one first name must not become one card. repair_duplicate_cards folds existing duplicates at boot on the same condition. link_person folds the card the speaker was on into the one they named, once no identity points at it any more.

When two cards do share a name, name lookups answer with the oldest.

Facts are capped per card (MAX_FACTS_STORED) and only the most recent are shown, so a card that keeps growing never eats the prompt.

Configuration#

"social_memory": { "enabled": true }

No other keys — the thresholds are constants in store.py and people.py (REGULAR_SESSION_THRESHOLD, MAX_FACTS_STORED, MAX_FACTS_SHOWN), because they are design decisions rather than settings.