nfrs.sgit.ai / memory

These sites are agent memory, done as publishing

The clearest statement of what this whole network is for was made in a chat message commissioning a site. It should not have stayed there.

“The idea of these sites is to provide good briefs for agents to learn about how we work and think — these sites are a more evolved and focused version of what is usually called LLM memory.”

This page develops that into the argument it deserves. It belongs on this site because the thesis is an NFR story: a memory is only better than a transcript if somebody maintains its non-functional requirements. It is also proposed for the network hub at sgit.ai, because it explains all fourteen sites at once rather than this one.

1. The claim

What products call “memory” today is an accumulation: embeddings of past conversations, retrieved by similarity, private to one vendor's silo, unversioned, uncurated, and invisible to the person it describes. This network is the same function built as publishing.

Conventional LLM memoryThese sites
Contentaccumulated transcriptscurated briefs — written, reviewed, pruned
Retrievalsimilarity search, opaqueaddressable URLs, plus llms.txt and llms-full.txt
Versioningnonegit, version-prefixed filenames, dated pages, a release history
Truthfulnesswhatever was saidthe reality-document rule: briefs are aspirations, not facts
Portabilitylocked to one vendorany agent, any vendor, one HTTP GET
LicenceunclearCC BY 4.0, explicit, irrevocable
Inspectable by the humanrarelyit is a website — read your own memory
Shared across agentsnoyes — one memory, many agents, no sync

Memory you can read, cite, version, license and hand to any agent — because it is a website.

2. The evidence that this was already the design

The thesis is not a retrofit. The estate was building toward it explicitly, before the sentence was said:

3. What “more evolved” concretely means

Five disciplines, and each is a non-functional requirement applied to knowledge rather than to code. This is the whole reason the page sits on this site.

1

Curation is the villagers' work, applied to knowledge

Somebody owns each site, prunes it, and keeps it true. That is the reality-doc discipline as memory hygiene, and it is maintenance work that produces no new features — which is exactly what makes it the villagers' job. Cross-link: open-source.sgit.ai owns the market-side argument.

2

Versioning makes memory correctable

The thing neither a frozen vault nor a transcript can do. A wrong page can be fixed, and the fix is itself in the record. See the release history.

3

The do-not-publish tiers are memory's security model

Every brief pack in this series ships one, and this site's is stated on the page that states its limits. A memory without redaction discipline leaks, and it leaks irreversibly, because publishing is not undoable.

4

Shipped-versus-asserted pages are memory's calibration

A memory that cannot distinguish designed from built teaches agents to overclaim. That is the single most common failure found across this estate's stale artefacts. This site's own: what it asserts versus what it can show.

5

Deconfliction is memory's normalisation

One canonical copy, siblings link. Without it the network grows diverging duplicates, which is what a hub is for and also what makes a hub dangerous. The contract: the topic map.

4. The honest limits

Staleness is the failure mode, and this estate has already demonstrated it four times. A capabilities file, a reality index left dozens of versions behind, a README describing a repository that does not exist, and a second version file read by nothing. In an ordinary documentation set those are untidy. In an agent-memory network a stale page is a false memory — which is why the thesis and the generate-or-date rule arrive together or not at all.

Curated memory is opinionated memory. These sites teach how we think, and an agent that reads them inherits the positions, including the wrong ones. The mitigation is the convention of publishing tensions unresolved rather than resolving them for tidiness, and it is a real mitigation rather than a disclaimer: the hub publishes the argument against its own existence.

Fourteen sites is itself a retrieval problem. Without the topic map and the cross-site deconflicts, the network reproduces at network scale exactly the discovery failure the agent-access report found inside one site.

Counter-evidence

The thesis is a claim about a property this network is trying to have, not a measured result. Nobody has run an evaluation asking whether an agent briefed from these sites works better than one that was not — and the scorecard's explainability row records that there are no evals anywhere in the estate. That absence applies to this page as much as to anything else on the site. The claim is falsifiable and untested, and it is published in that state deliberately.