Skip to content

How Memory Learns

See also:

AnySoul builds memory from continuous agent activity, not from one-off manual saves.

New memory usually starts from:

  • incoming events (messages, links, reminders, triggers),
  • focused reads during tool execution,
  • and outputs that produce stable reusable context.

The storage body is file-native, so memory can be listed, inspected, and reorganized through directory/file operations.

2) How Memory Is Reinforced or Weakened (Forgetting)

Section titled “2) How Memory Is Reinforced or Weakened (Forgetting)”

Memory strength is behavior-driven.

Reinforcement signals:

  • successful retrieval through search,
  • repeated co-occurrence in multi-step tool chains,
  • traversal through associative random-walk paths.

Weakening signals:

  • low activity over long horizons,
  • reduced contribution compared to newer or more useful relations.

Why forgetting exists:

Context windows are finite. Memory must prioritize what remains useful.
So low-value links tend to become quieter, while high-value links stay easier to activate.

3) Memory Backtracking, Cloning, and Mixing

Section titled “3) Memory Backtracking, Cloning, and Mixing”

Inspired by Git-style thinking, AnySoul explores identity-level memory operations:

  • Backtracking: each agent’s files are versioned so the agent can return to a specific day or timestamp.
  • Cloning: the current state of an agent can be copied into a new branch-like individual.
  • Mixing: selected memory segments from different agents can be combined in a controlled way.

This is our ongoing exploration of personality copying and personality mixing in agent systems.