Second Brain for AI
Persistent memory for Claude, ChatGPT & Cursor. Free.
Start new thread

How far should an AI follow your memory graph?

by
ā€¢

One of the hardest decisions in Second Brain v2 was deciding how much related context recall should include.

Imagine asking:

"What did we decide about the Product Hunt launch?"

A normal search might find the launch plan.

A graph-aware system could also find:

  • the positioning decision connected to it

  • feedback that changed the launch date

  • the demo you created afterward

  • a previous community post that performed well

  • an older decision that has since been replaced

That can be genuinely useful. It can also become noisy very quickly.

Second Brain handles this by ranking direct matches first, then expanding through connected memories with distance decay. The further away a memory is, the more evidence it needs to appear.

I want the graph to recover context you forgot existed, without making every answer feel like a tour of your entire database.

Where would you set the default?

  1. Direct matches only

  2. Direct matches plus one connection

  3. Follow several connections when confidence is high

  4. Let the AI decide each time

Second Brain v2 is live today, and this is one of the areas where I would especially value feedback.

202 views

Add a comment

Replies

Best

I like the distance decay idea. How will users adjust recall depth for different projects because one setting may not fit everyone?

Global setting for now. The MCP already exposes the hops parameter directly so you can tune it per call. Per-project presets are on the list.

me would start with one connected memory. users could expand further only when they need extra background and supporting details.

That's the current default. One hop keeps it tight without pulling in things you didn't ask for.

I think šŸ¤” SB v2 supposed to do exactly what you are doing here : ask us and let us to choose how deep and how far we want to go !

That's the direction. The hops parameter is already configurable per MCP call. Next step is surfacing it as a user-facing setting rather than a developer option.

Hi. At we are also concerned about the memory of our AI assistants. We chose to give the users full control on what is being remembered, and whether to reinforce the memories or prune them out.

Ā  User control is the right foundation. Second Brain keeps superseded memories linked in the graph rather than erasing them, so reinforce and prune are explicit choices with the history intact. Good to see building from the same premise.

It should be based one confidence.. so follow several connections if confidence is high, still keep it crisp.. as you mentioned distance is important only when there is confidence and context in the discussion which provides the evidence.. else most interaction should focus on current context only.. good wishes for your launch..

I would probably choose direct matches plus one connection as the default. It gives me enough background while keeping the answer quick and useful.

Ā That's the current default. One hop adds just enough related context without pulling in things two or three degrees out.

I like the distance decay idea. The challenge isn't finding more memories it's surfacing the ones that actually help answer the question.

Ā Exactly. Distance decay is specifically about precision, not volume. More recall only helps if what surfaces is actually relevant.

Context is valuable, but too much of it can become a distraction. I'd rather have a concise answer with the option to expand into deeper connections if needed.

Ā Direct matches always surface first, and the hops parameter lets you pull in connected context only when you need it. Tight by default, opt-in expansion when you want it.

option 3 feels right to me too, confidence-gated depth over a fixed hop count. one edge case I'm curious about - when the same memory is reachable through two separate connection paths (say, through both the launch decision and a person mentioned in both), does that count as reinforcing evidence and rank it higher, or does it just take whichever single path scored best and ignore the other? multiple independent paths converging on the same memory feels like a stronger relevance signal than distance alone would capture

Iā€™d probably default to direct matches plus one connection.

The value of memory is not showing everything related. It is surfacing the one forgotten context that changes the decision. If the graph expands too far by default, the user has to do the filtering again, which defeats the point.

12
Next