Portable Memory Delegation Loop
Combine persistent personal context with a current web task to reduce prompting
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 4
- Confidence
- 97%
The Portable Memory Delegation Loop combines persistent personal context with a short, current instruction. Stable information—family composition, location, preferences, or recurring constraints—lives in memory, while the user supplies only the immediate objective and changing details. The agent then searches or operates on the web using both sources of context and returns personalized results. This changes memory from passive chat history into a reusable input for browser actions. The loop is most effective when memories are accurate, relevant, and intentionally enabled. Users should review outputs for inappropriate assumptions and update stale context, especially where personal and work accounts remain separated. The result is shorter prompting, faster task initiation, and recommendations better aligned with the user's circumstances.
Origin
Extracted from Marketing Against The Grain during the hosts' discussion of Atlas bringing ChatGPT memory into browsing.
Core principles
- 01Persistent context should eliminate repeated background instructions
- 02Relevant memory improves agent personalization
- 03Current web content and long-term memory serve different roles
- 04Users should deliberately control which memories are available
How to run it
- 1
Establish stable memory
Record durable preferences, relationships, locations, and constraints that regularly shape decisions.
Pro tip Prioritize facts that would otherwise appear in many prompts.
Watch out Do not store unnecessary sensitive information.
- 2
Supply the current objective
Give the agent the immediate task and any time-sensitive details not covered by memory.
Pro tip Keep the instruction concise but include hard constraints.
Watch out Memory cannot replace task-specific goals.
- 3
Combine memory with the web
Allow the browser agent to use persistent context while gathering current options or completing the task.
Pro tip Ask the agent to explain which preferences influenced its choices.
Watch out Stale memory can personalize the result in the wrong direction.
- 4
Validate and update
Review the output, correct false assumptions, and update memory when stable circumstances change.
Pro tip Separate personal and work context where appropriate.
Watch out Account boundaries may prevent the agent from accessing all relevant memory.
In the wild
A user asks the agent to plan a vacation for specific dates. Existing memory supplies the spouse's name, children's names, and the family's preference for beach trips, allowing the agent to search for suitable destinations without a long briefing.
→ The agent produces family-relevant options from a much shorter instruction.
Common mistakes
Treating memory as always correct
Persistent context can become stale and should be reviewed when recommendations no longer fit.
Mixing incompatible identities
Personal and work memories may require separate contexts rather than one undifferentiated profile.
Is it for you?
Best for
Recurring planning, research, and recommendation tasks shaped by stable personal preferences.
Not ideal for
Sensitive tasks where persistent memory is unwanted or where preferences change frequently.
From the transcript
“I can just ask it to plan a vacation for a certain date, and it's going to give me options for our family and to…”
“Memory is huge. Like it personalizes the web for you in a way that was never possible with just kind of like history of Bryzen,…”
“And so you'll be able to bring your memory with you.”
From the episode
OpenAI Just Launched a Web Browser (and It’s Smarter Than Chrome)