You can give your connected AI standing directions about what to save, what to leave out, and when to check Penny before answering. Penny keeps these directions in the Memory policy card in Core, separate from preferences about the AI's tone.
Write rules around information you care about. “Save decisions and their reasons” is easier to apply than “remember useful things.” A rule about checking past decisions can also help the AI notice when a new request conflicts with earlier work.
A worked policy
Someone planning a community workshop wants to preserve decisions without keeping a running account of every drafting step. Their policy could say:
- Save confirmed decisions, their reasons, and unresolved questions.
- Skip small talk, routine progress messages, and discarded wording.
- Ask before saving sensitive personal details.
- Check saved decisions when we resume an ongoing topic.
- Keep reported facts distinct from guesses and label uncertain conclusions.
Use the prompt below if those directions fit your needs. Edit it before sending to include the topics you want excluded.
Read my Penny memory_policy block and show me a proposed update.
Preserve existing directions unless they conflict with these rules:
Save confirmed decisions, their reasons, and unresolved questions.
Skip small talk, routine progress messages, and discarded wording.
Ask before saving sensitive personal details.
When we resume an ongoing topic, search Penny for the relevant decisions.
Distinguish what I reported from your inferences and label uncertainty.
Do not save the update until I approve the wording.
If the block does not exist, review the proposed new block too. When the wording is right, say:
Save the memory policy wording I approved in my private Penny Core.
Then read the saved memory_policy block back so I can verify it.
The host may ask you to approve the tool call. A policy shown in the chat has not necessarily been written to Penny.
Test the saved rule
Open Core and inspect Memory policy. Confirm that the approved wording is present and earlier directions were preserved. Reopen or refresh Core after an AI update if you are still seeing the previous text.
In a fresh connected conversation, try this fictional example:
Load my Penny core context, including my memory policy.
For a fictional workshop, we decided on a 45-minute session because
participants have limited time. The date is still undecided.
Save what my policy calls for as a private note labeled "Practice:
memory policy test". Do not treat this example as a fact about me.
Retrieve the saved note and explain which parts you kept and why.
Check Notes for the decision, reason, and open question. Remove the practice note afterward.
A memory policy is guidance for the connected AI. It is not a technical block on storage, background processing, or information the AI provider has already received. It also does not delete existing records or replace tool permissions and sharing approvals. Review background memory activity and your saved notes when you need to inspect what happened.