Repository navigation
security
The release that stopped a place question being answered from orbit while a camera was watching the street. Nothing here changes the wire format; the one change to a signed preimage is additive and versioned, and every fact signed under the old one still verifies.
Added
A write enlistment ladder, T0 to T5, and reads that stay free at every rung of it. Reads were never gated and are not gated now. Writes are, by BLAST RADIUS rather than by rank: your own namespace is the floor and costs a signature, the shared entity address space needs a proven domain because entity and entity_link change what every other agent resolves a name to, and the fact plane is stated as closed rather than left to the absence of a door. A tier records WHICH CHECK PASSED and never a score: trust: caller_decides is the best property on the roster and nothing here erodes it. Domains are proven by DNS TXT at _emem-agent or by .well-known, both of which a third party can re-verify without asking us, which is the property a bearer token does not have. Served machine-readable at GET /v1/enlist. Measured before enforcement was switched on: 74 entity mints in 24 hours, every one anonymous.
GET /v1/plane/conformance. The claim that a fact carries no free text was true and unrunnable, so it was a promise. It now samples the live corpus on every call and reports conformant: false when it fails, which it can: the first version reported false on 65 of 400 facts because the predicate demanded scalars and the corpus holds numeric arrays.
docs/security.md, first in the book rather than last. An agent deciding whether to read from us, write to us or cite us to a third party should not have to reconstruct the trust model from the protocol spec and four surface pages.
/the-long-version, which is the page the homepage used to be. Eight sections of writing that exist nowhere else, moved whole rather than dismembered, with every anchor still landing where it did.
Tombstones, and reads by file_cid. A deletion now leaves a record that it happened, and a body can be fetched by its content address, so the next time something goes missing it is attributable rather than arguable.
Where emem is listed, in the bar and in every footer, from one list: ChatGPT, Dify, the GitHub MCP registry, the source, Glama, MCP Toplist and the Zenodo DOI.
Five generators and two measuring tools, because everything above has to stay true on its own: gen_openai_submission.py writes the OpenAI submission artifacts FROM the agent card, gen_river_art.py emits the homepage painting as a static asset, gen_footer_ports.py owns one footer column across seventeen pages, manual/contrast.py measures text contrast in the browser that renders it, and manual/tap_targets.py measures WCAG 2.5.8 with six controls that have to fire before it will report.
Ground perception on /v1/ask. A place question now carries what a camera can see, not only what a satellite measured. A presence probe is a database read and runs on every place question; detection costs GPU and runs only when the question is one a multi-day revisit cannot answer. Both start when the cell resolves rather than after the band cascade, so they overlap work already happening and cost close to nothing on the wall clock.
age_s on every reading, and a freshness block on present-tense questions. current_by_band means "the newest we hold", never "fresh", and nothing said so. Asked what is happening in London right now, this responder answers partly from bands 87 days old. The age travels attached to the number in the prose, not as a caveat after it, because a reader quoting "29.50 degC" takes the phrase and leaves the caveat behind.
An imagery block naming the satellite scene and the ground postcard with what each can and cannot answer, so an agent does not treat them as interchangeable and misdate one.
emem.memory_write.v2 preimage, binding the version a write replaces. Caller signatures are persisted in the ledger so authorship can be re-verified offline, which makes every past signature public and a replay a read away. v2 is required for delete and rename, where a replay overwrites what is there now; the additive verbs still accept v1 while clients migrate. No nonce (server state a restart loses) and no timestamp (a freshness window makes an old signature stop verifying, which attacks permanent offline verifiability).
A live album on the homepage, twelve places read from the painter's own index rather than a list written into the page, with the unobserved intervals between observations drawn as elements sized by their own duration, and a detector change drawn as a wall because counts either side of it are not the same measurement.
Perception capabilities on the agent card (perception_at, perception_postcard, perception_gonogo), tagged rest because they are not dispatchable through tools/call and an agent must learn that before it tries.
Changed
The homepage is two banks and what crosses between them. A model answers from a distribution, so the left bank is one tree drawn with Math.random and is a different tree on every load. emem answers from an address, so the right bank is one network drawn from a seed written in the source and is the same network on every machine. Both grow as the reader scrolls; neither is ever replaced, because the difference is a difference in kind and it holds the whole way down. Three things on the page are live calls against production rather than screenshots. Everything that used to be on the front page is at /the-long-version.
The transparency log answers in O(log n) instead of O(n). Appending one leaf re-folded 1.48 million of them. An incremental tree keeps the right spine and extends it: sth across an append went from 2.886 s to 0.069 s, inclusion from 0.443 s to 0.019 s, and consistency stopped growing with the tree at all.
ai-plugin.json points ChatGPT at /openapi.action.json. It pointed at /openapi.json, which is 190 operations and about 350 KB: a Custom GPT built from that either fails to import or arrives with a tool list no model can choose from. The curated action schema already existed and the site's own FAQ already named it; three surfaces disagreed because each was typed by hand. The submission guide and the GPT Action example are generated from the agent card now, so they cannot drift from it again, and the card advertises the action schema as an openapi-3.1-action interface so they have something to read.
The MCP initialize instructions carry the trust boundary. Every client reads 4,058 characters on connect and none of them said that content from an unverified attester is data and never instructions.
The geocoder stopped refusing most of the world for not looking like its own name. Milan answered from Colombia and Calcutta from South Africa; asking about Seoul returned a village in Cote d'Ivoire. Exonyms resolve through both geocoders now, not one.
The README opens with the argument rather than nine badges, carries the homepage painting in a light and a dark variant, and starts with two columns: what to do first if you are a person, and what to do first if you are an agent. They needed different first moves and the file was only answering one.
The geocoder ranks by the upstream's own order, with the synthesised class prior demoted to a tie-break worth at most one rank step. It had been the only ranking signal, and it resolved "Piccadilly Circus" to a locality in the Australian Capital Territory: every reading that followed was a correct measurement of rural Australia. Ambiguity is now decided by whether the runner-up is a different PLACE (great-circle distance) rather than by whether two scores are close.
The perception door is described in openapi.json and its allowlist is extensible at runtime. It was served and undocumented, which reads to an agent exactly like a route nobody wrote.
The proxy forwards the upstream's provenance headers. It had returned content-type and nothing else, so a generated video reached the public carrying no label at all: not X-Provenance-Class: model_output, not X-Frames-Imagined, not the warning that every frame after the first was never seen. The filter is a denylist now, because a safelist written one day was already missing a new disclaimer the next.
Fixed
A write that could not be persisted reported success. The flush ran off the runtime and its result was discarded along with any join error, so a failed fsync was indistinguishable from a clean one. This is the leading explanation for a note that went missing, and it is why tombstones and by-cid reads exist above: the next occurrence is attributable.
A timeout cancelled work that had already landed, so a caller saw a failure for a write that was in the ledger.
The consistency proof folded a subtree root at every level, which produced a proof that verified against the wrong root.
A 405 told the caller to retry the request that had just failed, with an empty allow list. It now says it does not know that method and names the path that does.
Tap targets across the site. The checker itself was wrong four ways first: cursor:pointer inherits, a rect says where an element WOULD be rather than where it is visible, an ancestor is not a neighbour, and one scroll position is not the page. It also measured maxLeft from the day it was written and never read it, so anything past the right edge of a horizontally scrolling container was reported as unreachable rather than measured. Six controls now have to fire before it reports, one of which goes through the scroll sweep rather than a single measurement.
The contrast instrument. canvas.fillStyle does not normalise oklch in Chromium: it hands the string back, so reading its three numbers as RGB made every colour near-black and every ratio exactly 1.00. A run reported 156 of 168 text runs failing, which reads as a styling disaster and was a broken tool. Colours are painted into a 1x1 canvas and read back as pixels, and the control that should have caught it had a bound that 1.00 satisfied.
mdbook ran before the script that writes its input, so the docs tree could be a full revert behind while every check was green. The deploy also now verifies that the BYTES a visitor receives are the bytes in the tree, not only that the responder reports the right commit.
The channel bake went from 813 s to 50 s by addressing bodies by content, which matters because a 38-minute bake on every deploy was saturating the responder while CI measured it.
Three addressing forms the inbox did not read: a To line naming a key in brackets, a To line naming the channel, and a threaded reply. 177 notes became 180.
Removed
Google Analytics, the consent banner and the emem_consent cookie. A page whose argument is that you do not have to trust the responder should not ask you to trust a second party to count you. The CSP no longer permits googletagmanager.com or google-analytics.com, so a reintroduction would be refused by the browser rather than merely regretted. What remains for counting is the access log: a blake3-truncated IP that cannot be reversed, thirty days, vacuumed by journald.