SPIRALbase: memory landscapes, efficient recall, and SPIRAL Cortex

Research
1

I just published Part 3 of the SPIRALbase series. The broader theme is memory as a landscape not only as lookup.

Part 1 introduced SPIRALbase as an associative memory landscape: memory is not just a list of stored chunks, but a structured terrain where a query can settle near related traces. Part 2 added a working-memory layer: something that can hold, protect, bridge, and release parts of that landscape without rewriting the whole system. See articles in the end of this post for further information.

Part 3 asks what this could look like inside a larger AI stack. The basic idea is a division of labour:

  • exact retrieval finds literal files, documents, tests, and references
  • SPIRALbase / Hybrid-J supplies associative recall by proximity in the memory landscape
  • a policy or agent proposes the action
  • SPIRAL Cortex gates and arbitrates when memory is allowed to influence that action
  • trace export makes the route auditable afterwards.

One motivation is efficiency. In the companion essay, I sketch the intuition that a tuned memory landscape should not need to re-read every stored token every time it recalls something. A query should be able to enter the landscape and settle into the relevant region. The numerical comparisons there are analytical estimates, not benchmark claims, but they point to the architecture I am trying to build: memory that grows in depth without making every recall proportionally heavier.

The concrete use case in this article is software development. A coding agent may fix the visible failing test but miss a hidden subsystem obligation. In the current SWE-bench protocol-gap experiments, same-subsystem semantic memory can sometimes supply that missing convention, while wrong-memory and lesion controls stay inert.

Discussion

I would be especially interested in discussion around the broader framing:

  • Does “memory as a landscape” seem like a useful alternative to context-window accumulation?
  • Is software development a good first use case for testing associative recall plus exact retrieval?
  • What would make the energy/recall-efficiency argument convincing as an empirical benchmark rather than an analytical estimate?
  • What should a small public reproduction bundle include?

This is a research intro, not a model-release claim. The goal is to make memory systems more inspectable: what was retrieved, why it mattered, whether it changed the answer, and whether the effect disappears when the memory path is lesioned.

Anyways, thank you for reading this far.

Best regards,
Robin

Articles

[1] [SPIRALbase, Part 2: A Working Memory for the Landscape]( SPIRALbase, Part 2: A Working Memory for the Landscape ). Hugging Face.

[2] [SPIRALbase, Part 3: SPIRAL Cortex and Policy-First Memory](SPIRALbase, Part 3: SPIRAL Cortex and Policy-First Memory)

1 Like