Skip to content
AnySoul Docs

Memory

AnySoul’s memory design is inspired by OpenClaw and one simple principle:

A capable agent should remember through a living workspace, not a frozen notes database.

Why File-System Memory

Section titled “Why File-System Memory”

We store memory as files because files are:

  • readable,
  • inspectable,
  • composable,
  • and naturally organized by directories.

This lets an agent understand context the same way humans do when opening a project folder:

  1. list directories,
  2. inspect candidate files,
  3. read deeper only when needed.

Why File-Native Memory Fits Proactive Agents

Section titled “Why File-Native Memory Fits Proactive Agents”

AnySoul treats memory as a file-system-first, NoSQL-like workspace:

  • files are the source of truth,
  • directory structure becomes semantic organization,
  • and list/read operations become native recall behavior.

This aligns with proactive agent behavior: agents can re-organize and revisit their own workspace while acting autonomously over time.

The Three Layers

Section titled “The Three Layers”

Memory in AnySoul can be understood as three layers:

  1. Storage: files and folders are the memory body.
  2. Retrieval: BM25 + vector search (plus grep-style matching intuition) activates relevant memory.
  3. Association: random walk + weighted links enables non-linear recall.

This produces a graph-like memory system: connected, uneven, and contextual.

Why Association Is Probabilistic

Section titled “Why Association Is Probabilistic”

Real memory is not a perfect lookup table.

Even if one memory has many possible neighbors, not all of them should appear in every recall. Context windows are limited, and relevance changes by task.

So we use weighted association with controlled randomness:

  • stronger links are easier to revisit,
  • weaker links can still surface,
  • but not every edge is expanded at once.

References

Section titled “References”
Section titled “Related”