Day 81: Long-Term Memory Without Becoming Creepy

Long-term memory makes agents feel smart until someone asks “why does it know my divorce?” Memory without consent, TTL, and a user-visible editor is just surveillance with embeddings.

⚡ TL;DR: Store memories as explicit, typed records with subject, source, consent, and expiry. Show them in UI. Default to session scratch + retrieval over silent profiles. Delete on request and on TTL.

Memory tiers (revisit Day 13 with ethics)

# ✅ Typed memory row — never a free-text void
from dataclasses import dataclass
from datetime import datetime

@dataclass
class Memory:
    id: str
    subject_id: str          # user or team
    kind: str                # preference | fact | procedure
    text: str
    source: str              # "user_asserted" | "inferred"
    consent: bool
    expires_at: datetime
    visible_to_user: bool = True

def write_memory(m: Memory, store) -> None:
    if m.source == "inferred" and not m.consent:
        raise PermissionError("inferred memory requires consent")
    if not m.visible_to_user:
        raise PermissionError("hidden memories forbidden")
    store.put(m)

❌ Silent inference of sensitive attributes from tickets or email into a vector profile.

User-visible controls

• List / edit / delete memories in settings.
• “Forget this project” wipes project-scoped rows.
• Export for DSAR.

Failure modes

“Helpful” inferred memories about health, finance, or HR topics. Block sensitive categories entirely. Embeddings of memory text can still leak via nearest-neighbor — encrypt vectors at rest and isolate per subject.

Closing checklist

• [ ] Consent flag and TTL on every long-term row
• [ ] UI to inspect and delete
• [ ] Inferred vs user-asserted distinguished
• [ ] No hidden memories
• [ ] Legal review of retention

Series navigation

Day 80: Project: Platform AI Gateway · Day 82: Voice and Realtime Agents

Last updated September 11, 2026