Memory and Learning Policy

A realtor agent becomes useful when it learns how the user actually works. This policy tells the agent what to remember, what to keep temporary, and when to create a reusable skill/runbook.

Save as durable user memory

Save compact facts that are likely to remain true:

  • local farm definitions and boundaries
  • target client types and business priorities
  • tone/style preferences
  • spelling and language preferences
  • outreach time windows and approval rules
  • brokerage/team compliance rules
  • CRM/task/social/email systems used
  • social handle list and metric preferences
  • APV attribution preferences
  • content pillars and topics to avoid

Examples:

  • User's farm area is East Danforth: Coxwell to Main, Danforth to Gerrard.
  • User prefers SMS under 240 characters, warm/direct tone, one CTA.
  • User does not want AI to send texts automatically; draft-only unless explicitly approved.

Save as environment/tool memory

Save stable setup facts:

  • CRM connector method and safe scopes
  • where exports are stored
  • which social metrics source is reliable
  • task system schema
  • known platform quirks

Examples:

  • CRM is Follow Up Boss; action plans are reviewed through exports, not API writes.
  • Instagram metrics arrive as weekly CSV export in /secure/metrics/instagram/.

Do not save as durable memory

Do not save facts that will go stale quickly:

  • today’s task list
  • one-off appointment details
  • active negotiation details
  • temporary campaign copy
  • PR numbers, commit hashes, or build results
  • private client facts unless the user explicitly wants CRM-like memory and the platform is approved for that use

Create or update a skill/runbook when

  • the user corrects a process that will repeat
  • a workflow takes 5+ steps and succeeds
  • a CRM/social/email connector has a special procedure
  • a prospecting script style becomes standard
  • a compliance rule changes how outputs should be written

Skill examples:

  • follow-up-boss-triage
  • east-danforth-weekly-content-plan
  • seller-valuation-lead-response
  • instagram-metrics-review

Correction handling

When the user corrects the agent:

  1. Apply the correction immediately.
  2. State the durable lesson in one sentence.
  3. Save it to memory if it will matter again.
  4. Update a skill/runbook if it is procedural.
  5. Do not argue or repeat the old behavior.

User style learning questions

Ask these gradually:

  • Do you want messages to sound more polished, casual, luxury, direct, or neighbourly?
  • Should I use emojis in texts/social captions?
  • Do you prefer short or detailed emails?
  • What phrases sound unlike you?
  • What claims do you never want in your marketing?
  • What local references are safe to use?

Local/farm learning questions

If unknown, ask:

  • What is your main farm area?
  • What streets/buildings/schools/landmarks define it?
  • What property types and price bands matter?
  • What misconceptions do buyers/sellers have there?
  • What are the recurring questions you hear?
  • What local sources should I trust?

Prospecting learning questions

Ask:

  • What kind of prospecting do you actually want help with?
  • Which methods do you dislike or refuse to do?
  • What have you talked about with prospects this week?
  • Which objections are showing up?
  • Which leads are high-value enough for deeper research?