The memory thesis
Summary
Develops one sentence from the commissioning message into an argument. Conventional LLM memory is an accumulation of transcripts, retrieved by similarity, private to one vendor, unversioned, uncurated and invisible to the person it describes. This network is the same function built as publishing: curated briefs, addressable URLs, git versioning, an explicit licence, and a human who can read their own memory. The document's strongest section is the evidence that this was already the design before anybody said it: guides with for_llms in the filename, the markdown twin at every URL, and an agent-access report that diagnosed a memory-retrieval failure in exactly those terms. It closes with three honest limits, of which staleness is the sharpest: in an agent-memory network, a stale page is a false memory.
Key concepts
- The thesis page
- The evidence it was already the design
- Staleness as the failure mode
- The agent surface
Key ideas
- Memory you can read, cite, version, license and hand to any agent, because it is a website.
- Curated memory is opinionated memory: an agent reading these sites inherits the positions, including the wrong ones.
- Fourteen sites is itself a retrieval problem, which is what the topic map is for.
- The do-not-publish tiers are memory's security model; shipped-versus-asserted pages are its calibration.
On this site
Published as /memory/, and recommended to the network hub at sgit.ai, where it would explain all fourteen sites rather than one. That move is listed as proposed and not built.