Changelog
Each release of Retiform's MCP tools and REST API is listed here, newest first. A change that needs something from you says what to do and by when.
The trial accepts PDF files.
- You can upload a PDF through the console's Documents page, the kg command or the REST API. Each page is turned into an image and read by a model, so a PDF takes longer to become searchable than other files.
- A PDF file can be up to 20 MiB and 300 pages, and a tenant's PDFs are read up to 1,000 pages a month. GET /limits now lists that allowance as document-pages, and the console's Usage page shows it.
- Retiform keeps the PDF as you uploaded it and the text of each page. It keeps no page images.
- The REST contract and the MCP server are unchanged.
A memory entry can say that it settles an earlier one, such as a fact that settles a recommendation, so recall ranks the fact first while the recommendation stays readable.
- record_memory, supersede_memory and their REST operations take settles: up to 5 current entries of your tenant that the new entry answers or completes. A settled entry stays current, with its statement and history unchanged; recall and context never rank it above an entry that settles it on the same page.
- Every memory entry shows settles, the entries it settles, and settledBy, the current entries that settle it, newest first. Correcting an entry without settles carries its links forward, less any entry no longer current; an empty list clears them. Correcting a settled entry starts unsettled.
- A write naming an entry that is superseded, is the entry being superseded, or is not a current entry of your tenant is refused with MEMORY_INVALID, naming each one; an entry of another tenant reads the same as one that does not exist. A write with settles made while the tenant's schema is being upgraded is refused with MEMORY_SETTLES_UNAVAILABLE and can be retried, or sent without settles.
- The REST contract is 0.69.0 and the MCP server is 0.36.0. record_memory and supersede_memory take settles, and every tool that returns memory entries gains settles and settledBy; the other tools are unchanged.
Search tells you when nothing it found is relevant, and how each result matched.
- Each search result now carries vectorSimilarity, its cosine similarity to the query rounded to 4 decimals (null when only the keyword list found it), and matchStrength: all-terms, partial, few-terms or semantic. Unlike fusedScore, the similarity can be compared across queries.
- Each page carries nothingRelevant: true when no result on it passes the relevance rule assemble_context and recall_memory already use. On a hybrid first page it also follows context's abstention check, so search and context agree for the same question. Results, their order and cursors are unchanged; no result is removed.
- The REST contract is 0.68.0 and the MCP server is 0.35.0. The MCP search tool's output schema and description name the new fields.
The MCP server asks connected agents for a hand-off right after their context is compacted, not before.
- The instructions in the initialize answer now place the fourth moment, the session hand-off, at the end of a session and right after a compaction, with supersede_memory to replace the session's previous hand-off. An agent cannot tell when a compaction is coming, so asking before it was rarely followed.
- The REST contract is unchanged and the MCP server is 0.34.0. The MCP tools are unchanged.
The usage page and the context-usage report show the size of your imported instruction files beside a typical recall answer.
- GET /me/context-usage carries imported, with documents and tokens: how many instruction files the Claude Code plugin's import uploaded are stored now, and the tokens in their current revisions. It is absent, never zero, for a team that has imported none.
- The files are the documents the import named after one of your projects. They are part of the stored figure, follow each new revision and deletion, and are read at most once a minute.
- The console's usage page shows the figure beside the typical recall answer of the chosen period, which is your whole team's, or one line saying how to make it appear.
- The REST contract is 0.67.0 and the MCP server is 0.33.0. The MCP tools are unchanged.
The limits you see list a PDF page allowance only once PDF uploads are switched on.
- GET /limits leaves out document-pages while this deployment does not accept PDF files, so the console and Usage no longer suggest that PDF uploads work during the trial. The plan's allowance is unchanged and appears as soon as PDF is switched on.
- The REST contract is 0.66.0 and the MCP server is unchanged. No schema changes.
The console's Usage page splits this month's retrieval calls into searches, context requests and memory recalls, and says how many were not answered.
- GET /me/context-usage gives retrievals: this calendar UTC month's retrieval calls charged against the plan, by kind (search, context, recall), and failed, the context and recall calls charged but not answered. It is null where the platform schema is absent.
- The REST contract is 0.65.0 and the MCP server is 0.33.0. The MCP tools are unchanged.
The memory page can list everyone who wrote entries, and you can withdraw your own review of an entry.
- GET /memory/authors lists every person and agent who wrote a current memory entry, with how many each wrote and when last. An agent carries its names and the session label it declared. A person is marked when they are you, and their email address is included only for an administrator of the tenant.
- DELETE /memory/entries/{entryId}/reviews/mine withdraws your own review of an entry; other people's reviews stay.
- The REST contract is 0.64.0 and the MCP server is 0.33.0. The MCP tools are unchanged.
Administrators can withdraw an outstanding invitation, or send it again.
- DELETE /onboarding/v1/tenant/invitations/{invitationId} withdraws an invitation that has not been accepted: its link stops working and its place no longer counts against the member limit.
- POST /onboarding/v1/tenant/invitations/{invitationId}/resend invites the same address with the same role again, replacing the old invitation with a new expiry. Like any invitation, it is answered pending whatever the address.
- The REST contract is 0.63.0 and the MCP server is 0.33.0. The MCP tools are unchanged.
Recording a memory entry names recent entries that say nearly the same, so you can correct one instead of keeping two.
- A recorded entry carries possibleDuplicates: the current entries among the 50 newest that share at least half their words with the new statement, most similar first, at most three. The entry is stored either way; when one states the same thing, supersede it.
- The REST contract is 0.62.0 and the MCP server is 0.33.0. record_memory's result gains possibleDuplicates; the other tools are unchanged.
An invitation's accept page can show which tenant you are joining, with what role and for which address, before you join.
- GET /onboarding/v1/invitations/{invitationId} returns an open invitation's tenant name, role, address and expiry, only to the signed-in person it was sent to. Anyone else gets the same answer as for an invitation that does not exist.
- The REST contract is 0.61.0 and the MCP server is 0.30.0. The MCP tools are unchanged.
The console's agent list says which session each agent last worked in.
- GET /onboarding/v1/tenant/agent-client-approvals gives each agent its lastSessionLabel: the session label it declared with its latest memory write in the last 30 days, or null.
- The REST contract is 0.60.0 and the MCP server is 0.30.0. The MCP tools are unchanged.
The console's account page lists your kg command-line sign-ins, and you can end any of them.
- Account now shows each sign-in the kg command holds for you, for uploading or for reading, with when it started, when it was last used, when it expires and the last program that used it.
- End one and the command's next renewal is refused, so that copy of kg must sign in again. The command's own sign-ins cannot list or end sign-ins: only the console can.
- The REST contract is 0.59.0 and the MCP server is 0.30.0. The MCP tools are unchanged.
Context assembly can say nothing is relevant when a question asks for a version, an amount or a named region that the records don't hold.
- Contexts now report rules context-v4, and configuration.abstention carries answerType: off, not-applicable, found, missing or not-evaluated, with the kind of value asked (version, amount or named-region) when one is recognised.
- With the check on, a question asking for one of those kinds whose passages hold no value of that kind answers nothingRelevant instead of passages about the topic. It ships switched off.
- The REST contract is 0.58.0 and the MCP server is 0.30.0. The MCP tools are unchanged, so a connected agent sees the new field without a restart.
A memory entry's author shows the name its agent's approving person gave it.
- A memory entry's attribution carries tenantData.clientApprovedName: the name the person who approved the agent client gave it (#1095), read at each request, so renaming the agent shows on the entries it wrote earlier. It is null when the person gave no name, and for a person's own entry.
- The name is a member's own text, so it sits under tenantData, which results mark as data, never instructions. Prefer it to clientRegisteredName, the name the client gave itself.
- The REST contract is 0.57.0 and the MCP server is 0.29.0. An agent connected before this release keeps its old tool list: restart it to see the new field.
The context-usage report says how often each agent used Retiform: on how many days it was active, read, wrote, and read before writing.
- GET /me/context-usage carries agentDays: for the last 7 and 30 UTC days, each agent's days (active, withRead, withWrite, withNeither, readBeforeWrite and readBeforeWriteUnknown) and the tenant's sums, so an agent that connects and never reads is listed with its active days and zeroes.
- A read is an answered context or recall request; a write is a memory entry recorded or superseded. Writes, and the writes per session label a client declared, are read from the tenant's own audit trail for each report; when it cannot be read, writesRead is false and those figures are null, never zero.
- The REST contract is 0.56.0; the MCP tools are unchanged.
The memory list filters by author, a date and text, and the new-entries list says how many are new.
- GET /memory/entries takes since (a date or a date-time, as recall's), author (an entry's attribution.principalId), authorKind (human or agent) and contains (text the statement holds, case-insensitively, matched as typed). They combine with each other and with labels, newest first, and none counts as a retrieval call.
- GET /memory/entries/new carries newCount, how many entries are new in all, exact up to 1,000, and newCountComplete, false when there are more.
- The REST contract is 0.55.0; the MCP tools are unchanged.
The documents list's state filter finds exactly the documents reporting each state, embedding included.
- GET /documents' state filter accepts embedding: a committed revision whose passages are still being embedded for meaning search, as latestRevision and GET /document-jobs/{jobId} report it.
- state=succeeded no longer matches a document reported embedding, so a document's reported state and the filter that finds it always agree. Ask state=embedding,succeeded for every committed document.
- The REST contract is 0.54.0. The MCP server is unchanged.
Context says whether hybrid coverage is on.
- assemble_context's and POST /context's configuration now carries hybridCoverage, on or off: whether hybrid search completes its keyword list with the documents that hold most of the question's words. It is off by default. REST API 0.53.0.
Equally scored search results now list the newer record first.
- When two results of search, context or memory recall score exactly the same, the one created later now comes first, then the identifier as before. Ties are common in hybrid search: a record only the keyword list found and one only the vector list found at the same rank score the same. Which records a list holds is unchanged.
- A search cursor issued before this release is refused with SEARCH_CURSOR_INVALID and the hint to restart without a cursor; restart the search to continue. REST API 0.52.0.
The MCP server tells a connected agent when to use it, and results name the next step.
- The initialize answer now carries instructions: at the start of a task, recall the hand-off; before acting on something the team may have settled, ask for context; after a decision, record it; at the end of a session, leave a hand-off. They reach the model even when a client defers tool descriptions or runs tools from code.
- assemble_context, search, read_passage, recall_memory, record_memory, supersede_memory and upload_document add one line to their result's text, after the tenant line, naming the sensible next step. structuredContent is unchanged, and the content stays the text's last line.
- The MCP server is 0.28.0; the REST contract is unchanged. An agent connected before this release keeps its old session: reconnect it to receive the instructions.
The memory list filters by label, and an agent's entries name the client it registered as.
- GET /memory/entries takes labels (repeated or comma-separated): only entries carrying every one of them are listed, newest first, as recall's labels filter keeps them.
- A memory entry's attribution carries clientRegisteredName: for an agent's write, the name its client registered with. The client chooses that name itself and it is not verified, so show it only beside the client identifier and marked as self-reported. It is null for a person or a client the platform no longer knows.
- The REST contract is 0.50.0 and the MCP server 0.27.0. An agent connected before this release keeps its old tool list: restart it to see the new attribution field.
The documents list says each document's state, type and size, can count, search, filter and order, and a revision's passages can be read with their text.
- Each item of GET /documents now carries latestRevision: its newest revision's processing state, exactly as GET /document-jobs/{jobId} reports it (queued, running, embedding, succeeded, failed or cancelled), with the failure reason, the media type and the size in characters.
- GET /documents?count=true adds total and totalExact: how many documents the list holds, exact up to 10,000.
- GET /documents can search and filter: contains (part of the title or path, case-insensitive), prefix (the start of the path), state (the newest revision's job states, comma-separated) and order (created, the default, or changed, title or externalId). The total counts under the same filters, and a cursor is refused under a different search, filter or order.
- GET /me/limits' response schema is now named PlanLimits. It shared the name Document with the documents API, so the reference showed documents with the plan limits' shape. No response changes.
- GET /documents/{documentId}/revisions/{revision}/passages lists a revision's passages with their text, up to 100 a page from fromOrdinal, with nextOrdinal when more follow.
- The REST contract is 0.49.0. The MCP server is unchanged.
Documents can be found by title or path with search.
- A document's title and external id (the path it was uploaded under) are now searchable, so search with types ["Document"] finds "the 007 plan" instead of being refused, and a search without types can rank a document beside its passages. Context and memory recall are unchanged. Existing tenants get this from the starter schema's version 5, applied by the operator run after the deploy; each tenant's documents are indexed in the background at the platform's cost, not from the tenant's embedding allowance. A tenant that had turned search off for these properties itself has it turned on again and can turn it off with setPropertySearch.
Extraction's provider refusal and output errors now reach the client with their codes.
- POST /extractions/preview answers PROVIDER_REFUSAL and PROVIDER_OUTPUT_TOO_LARGE with 422 (change or split the text) and PROVIDER_MALFORMED_OUTPUT and PROVIDER_CONFIGURATION with 500, instead of 502. A 502 was replaced at the edge by a generic 503, so a client never saw these codes and could mistake them for a passing outage. The codes and their meaning are unchanged, and none is retryable as it is. REST API 0.48.0.
Every MCP tool result names the tenant it came from.
- A successful tool result carries structuredContent.tenant: the tenantId, and the tenant's display name under tenantData.displayName (the name is the tenant's own text). A tenant without a display name is named by its tenantId. Error results are unchanged.
- The result's text gains one line between the notice and the content, Tenant: <tenantId> (tenantData.displayName: "<display name>"). The name is labelled as tenant data, which the notice covers, and quoted as a JSON string. The content is still the text's last line.
- Each tool's output schema lists the optional tenant object. The MCP server is 0.26.0; the REST contract is unchanged. An agent connected before this release keeps its old tool list: restart it to see the new field.
Context's omitted list says what each left-out item is.
- Each item of assemble_context's and POST /context's omitted list now carries its entityType, a title (a passage's document title, or an entity's title, name or memory statement) and a passage's headingPath, each cut to 120 characters, so an agent can judge whether to raise the budget without reading each item. REST API 0.47.0.
Context assembly now says nothing is relevant when both its vector and keyword evidence are weak.
- The context-v3 abstention gate is on by default. A context whose best vector similarity is below 0.45 and where no passage holds two thirds of the question's content words answers nothingRelevant instead of unrelated passages; configuration.abstention says answered or abstained.
- Measured on the development questions: unsupported questions answered nothingRelevant rose from 5 to 10 of 14, and 2 of 66 supported questions now answer nothingRelevant too. Questions that name a real topic but ask for a fact the records lack can still get passages.
- No contract change: the REST contract stays 0.45.0 and the MCP server 0.24.0. An operator can turn the gate off with KG_CONTEXT_ABSTAIN_ENABLED=false.
A memory recall can ask only for entries written since a date.
- recall_memory and POST /memory/recall take an optional since: a date such as 2026-11-18 (from 00:00 UTC) or a date-time with an offset such as 2026-11-18T06:00:00Z. Only entries written at or after it are returned, by the time attribution.recordedAt reports, so a correction made today counts from today.
- since works alone (the newest entries since then), with labels (for example the hand-offs since yesterday) and with a topic, which still ranks. A cursor is bound to it, and a since of another shape is refused naming it.
- The REST contract is 0.46.0 and the MCP server 0.25.0. An agent connected before this release keeps its old tool list: restart it to see since.
Context assembly can say nothing is relevant when both its vector and keyword evidence are weak.
- Contexts now report rules context-v3 and a configuration.abstention object: off, answered, abstained, not-needed or not-evaluated, with the similarity and coverage thresholds.
- With the gate on, a context whose best vector similarity is below 0.45 and where no passage holds two thirds of the question's content words answers nothingRelevant instead of unrelated passages. It ships switched off; it is switched on once its development-set measurement holds.
- The REST contract is 0.45.0 and the MCP server is 0.24.0. The MCP tools are unchanged, so a connected agent sees the new field without a restart.
Search no longer returns document passages that hold only a heading.
- search and POST /entities/search leave out a passage whose text is only headings, such as "## Assumptions", which matched only through its document's title in the heading path. Context already leaves these out. A page reads on past them, and past an entity deleted while the page was read, so it holds pageSize results unless the results end.
A malformed Query DSL filter or criteria now says which part is wrong and what it accepts.
- search (filter) and query_entities (criteria) name the failing place and what it accepts, for example "filter.kind: 'comparison' is not one of predicate, and, or, not" or "criteria.field: an object with scope (core or property) and name". POST /entities/search returns the same message.
- The filter and criteria inputs describe a predicate with a worked example, the operators, and how to combine expressions.
Keyword search finds questions written in plain words, and the generic tools explain how memory-entry labels are stored.
- Under the simple analysis, keyword search no longer requires a query's unquoted English stop words, so "how do I recover a stale lock" requires recover, stale and lock. Quoted phrases, exclusions and OR keep their words, and a query of stop words alone is kept as written. This applies to search, context and memory recall.
- The search, get_entity, query_entities and get_neighborhood descriptions now say that a memory entry's labels hold their stored form ",a,b," (filter with contains ",a,") and that recall_memory, which returns labels as a list, is the way to read entries.
Context assembly can rerank its candidates, so a passage that answers one part of a two-part question can be chosen.
- Contexts now report rules context-v2 and a configuration.reranking object: status applied, or skipped with a reason (off, request, unavailable, time-limit or refused), plus the reranker, depth and time limit. A skipped context is in the same fused order as before.
- The reranking stage splits a multi-part question at its clause boundaries, searches each part and merges the lists. It calls no model and stores nothing. It ships switched off; it is switched on once its development-set measurement holds.
- POST /context and assemble_context accept reranking: "off" to ask for the fused order.
- The REST contract is 0.44.0 and the MCP server is 0.23.0. An agent connected before this release keeps its old tool list: restart it to see the new option.
An upload's own sign-in can follow the processing of the documents it sent.
- GET /document-jobs/{jobId} now also answers a caller holding only kg:write, such as the kg command's sign-in, for a job that caller requested. Any other job, existing or not, is still 403 INSUFFICIENT_SCOPE, and documents, passages and search still need kg:read.
- The REST contract is 0.43.0. The MCP server is unchanged.
Context says how many items filled unused budget.
- The context configuration now has filled: how many items with role "fill" the response holds, 0 when none. A reader can tell that fill below the relevance cut-off was used without checking every item's role.
- The REST contract is 0.42.0. The MCP tools are unchanged, so a connected agent sees the new field without a restart.
A memory recall by topic now fills its page instead of returning a short one with more to come.
- recall_memory with a topic keeps only entries relevant to it. It used to return one page of candidates after dropping the rest, so asking for 8 could return 2 with hasMore true. It now reads further candidate pages, up to five per call, until the page is full or the candidates end.
- A page that is still short while hasMore is true says why in guidance, and the tool description says a page can be short. Nothing is skipped: the next page continues where this one stopped.
- The MCP server is 0.22.0; the REST contract is unchanged. An agent connected before this release keeps its old tool list: restart it to see the new description.
Deleting an entity that memory entries refer to no longer removes their provenance silently.
- When an entity a memory entry is about or cites is deleted with cascade=true, the entry keeps its statement and attribution and lists the removed link in a new deletedLinks field, as {type, entityId, deleted: true}. Only the link type and the deleted entity's id are kept, never its name or properties.
- An entity that a superseded chain of entries refers to can now be deleted this way; before, the links went without a trace.
- The REST contract is 0.41.0 and the MCP server is 0.21.0. An agent connected before this release keeps its old tool list: restart it to see the new field.
Context can use budget it left unused for one more passage just below its relevance cut-off.
- When the relevant passages leave part of the token budget unused, assemble_context and POST /context may add one more search result that fell just below the relevance cut-off. It is marked role "fill", keeps its citation, and is weaker evidence than the others. A question nothing in the records answers still says nothing is relevant.
- The configuration now echoes fillBelowFloor (1 by default; operators set it with KG_CONTEXT_FILL_BELOW_FLOOR, and 0 turns it off).
- The REST contract is 0.40.0 and the MCP server is 0.20.0. An agent connected before this release keeps its old tool list: restart it to see the new role.
A client can name the session that wrote a memory entry, as a declared, unverified label.
- A memory write over MCP or REST may carry a KG-Session-Label header: 1 to 40 characters of letters, digits and . _ : -. The entry's attribution then shows it as declaredSessionLabel, so several sessions that share one agent can be told apart. A value outside that bound is ignored and the write goes ahead.
- The label is what the client declared, not something Retiform verified: two sessions may declare the same one, and nothing uses it for access, limits or recall.
- The REST contract is 0.39.0 and the MCP server is 0.19.0. An agent connected before this release keeps its old tool list: restart it to see the new attribution field.
Agents can create an entity over MCP, such as the project a memory entry is about.
- The new create_entity tool creates an entity of a type in your schema and returns its id, so an agent in a new tenant can create the Project, Organization, Person or Decision that record_memory's about points at. It needs kg:write and checks the properties as POST /entities does.
- If an entity of that type already has the identifying value given, create_entity returns it unchanged with created false, so an agent can call it to make sure a subject exists. Its stored properties come back only to a caller that also holds kg:read; a write-only caller gets its id.
- The MCP server is version 0.18.0. An agent connected before this release keeps its old tool list: restart it to see create_entity.
kg upload and the console title a Word file by its first heading.
- A Word (.docx) file is titled by its first Heading 1, else by its document title property, else by its path, as Markdown and HTML files are titled by their first top-level heading. Uploading an unchanged Word file again gives an earlier upload the new title.
A document preview's duplicate suggestions compare every entity of the type, and say when they could not.
- Each entity candidate's possibleDuplicates in a document preview now compares every entity of the type. Before, it compared only the type's oldest 10,000, so a tenant with more entities of a type got no suggestion from the newer ones.
- While entities stored before this release are still being indexed for the comparison, a candidate that was looked up carries possibleDuplicatesComplete: false. The field is absent once every entity can be compared.
- The REST contract is 0.38.0.
Uploading an unchanged document with a new title updates its title.
- An upload whose content equals the document's current revision still answers unchanged, with no new revision or job, but a different title now replaces the document's title. kg upload and the console derive better titles than before (the first heading, or the path), and documents uploaded earlier take them at their next upload.
Releases hand over between two instances, and a request for leader-only work during the handover is answered 503 with Retry-After.
- During a release, a new instance starts beside the old one and takes over its single-instance work. For the second or two of the handover, a request for that work (starting an audit verification or export, a tenant deletion or its cancellation, an erasure confirmation or a person's own erasure request, and approving or revoking agent clients, service applications, invitations and member changes) may be answered 503 SERVICE_BUSY with retryAfterSeconds and the Retry-After header. Retry after the stated seconds. Every other request is served throughout.
- The REST reference documents this refusal as LeaderBusy and in the 503 descriptions of the routes it applies to.
- The REST contract is 0.36.0.
An upload job reports embedding while its passages are still being embedded.
- get_document_job and GET /document-jobs/{jobId} report the state embedding once a job's passages are stored and keyword search finds them, while vector embeddings are still being made. The state becomes succeeded once every passage is embedded, or when embedding waits for a reason that embeddingPending names (no provider configured, or the allowance spent). A client that waits for succeeded, failed or cancelled now waits for vector search too.
- The REST contract is 0.35.0.
A memory recall by topic returns only entries relevant to it, and says when none are.
- recall_memory and POST /memory/recall with a topic return only entries relevant to it: ones that match the topic's words, or whose meaning matches closely enough (the same floor assemble_context uses). Before, a topic returned the best-ranked entries even when none was about it.
- Each entry a topic recall returns has matchStrength: all-terms, partial, few-terms, or semantic when only its meaning matched. The result has nothingRelevant, true with a guidance line when no entry on the page is relevant. With hasMore false, that means there is no earlier conclusion on the topic; with hasMore true, a later page may still hold one. A page may now hold fewer entries than pageSize.
- A recall without a topic (about or labels alone, newest first) is unchanged.
Model usage shows when the deployment's shared embedding capacity is spent.
- GET /model-usage's allowance has deploymentCeilingReached. It is true when the monthly embedding capacity that every tenant shares is spent: embedding is then refused with ALLOWANCE_EXHAUSTED until allowance.resetsAt, and search answers keyword-only, even though your own allowance has room. It is a flag only and never shows other tenants' usage.
- The REST reference now shows the model allowance's own fields (ModelAllowance) instead of the request rate limit's.
File memory entries under labels, and recall them by label.
- record_memory and supersede_memory, and their REST routes, take labels: up to 10, each 1 to 32 lowercase letters, digits and . _ : / -, such as handoff or decision. A label of another shape is refused naming it. An entry keeps its labels sorted and never changes them; to relabel, supersede it. A supersession without labels keeps the old entry's, and an empty list clears them.
- recall_memory and POST /memory/recall take labels and return only entries carrying all of them, alone or with a topic or about. Labels alone list the entries newest first, so labels ["handoff"] is the work left for the next session. Labels are not searchable.
- Every memory entry shows its labels, empty when it has none. A label holding a credential is refused with SECRET_DETECTED. A write with labels made while the tenant's memory type is being upgraded is refused with MEMORY_LABELS_UNAVAILABLE and can be retried, or sent without labels.
Entities, bulk imports and extraction text that contain a credential are refused.
- POST /entities and PATCH /entities/{entityId} refuse a property whose text contains a credential of a listed kind with SECRET_DETECTED, naming the property, the kind and its line and column, and store nothing. An update checks only the properties it sends.
- In POST /entities/bulk, an item holding a credential is rejected with a secret object naming the same; the other items are written.
- POST /extractions/preview refuses such text before anything is sent to the model provider. A document section's preview refuses a section holding one (only a revision stored before this check can) with the document error SECRET_DETECTED, located in the revision. Text without a credential is unaffected.
Memory entries and documents that contain a credential are refused.
- record_memory, supersede_memory and upload_document, and their REST routes, refuse text that contains a credential of a listed kind with SECRET_DETECTED, naming the kind and its line and column, and store nothing.
- A PDF's processing job fails with the same code and location when its read text contains one.
- Text without a credential is unaffected. See Credentials in what you write for the kinds found.
Review what your agents recorded since you last looked, and accept or correct it.
- GET /memory/entries/new lists the current memory entries recorded since you last marked them seen, newest first; POST /memory/entries/new/seen moves your marker. Both are a person's: an agent's token is MEMORY_MEMBERS_ONLY. Entries recorded in the last minute stay listed until the next time.
- POST /memory/entries/{entryId}/reviews records that you accepted an entry as it stands. Every memory entry now carries reviews, newest first, in the REST responses and in the MCP memory tools' results. Requests that were valid before are unchanged.
The memory list can show only entries whose source changed.
- GET /memory/entries takes sourceChanged: include (the default), only or exclude. With only or exclude, one request examines at most 500 entries; boundReached says when that, not the end of the list, ended the page, and nextCursor continues the search.
- The console's memory page marks entries whose source changed, can list only those, and records a check that an entry still holds. Requests that were valid before are unchanged.
Memory entries say when a cited document changed since they were written.
- Every memory entry now carries sourceChanges: each cited passage whose document has a later revision (revised, with the current revision) or was deleted. assemble_context's memory items carry it too. It says the source changed, not that the entry is wrong.
- confirm_memory and POST /memory/entries/{entryId}/confirmations record that an entry was checked against a cited document's current revision, or its deletion, which clears the mark until the document changes again. The statement and who wrote it stay unchanged.
- recall_memory and POST /memory/recall take sourceChanged: include (the default), only or exclude. Requests that were valid before are unchanged.
A busy search is refused at once, with the time to retry.
- When search capacity is in use, POST /entities/search answers 503 SERVICE_BUSY straight away, with a Retry-After header and retryAfterSeconds in the body. Before, it could wait up to a second first.
- The MCP search tool answers UNAVAILABLE in the same case, retryable after the same time.
- Capacity is shared fairly between tenants, so one tenant's heavy searching does not lock out another's. Requests that were valid before are unchanged.
A new read shows what your agents were given, beside the size of what you store.
- A new operation, getContextUsage (GET /me/context-usage), reports for today, the last 7 days and the last 30 days how many context and memory recall requests were answered and refused at the monthly retrieval limit, with the tokens served in total and the typical answer size, for your team and for each agent.
- It also reports the tokens in your documents' current revisions and current memory entries at the time of reading. Sizes use the kg-utf8-quarter-v1 counter, Retiform's own estimate.
- The console's usage page shows the same figures. Nothing changes for current callers.
A bulk entity import says whether its rows are searchable yet.
- POST /entities/bulk answers with searchIndexing: immediate when the imported rows are searchable at once, or deferred when a large import leaves their keyword documents to a background indexer.
- While rows are deferred, search coverage counts them as not yet indexed; they become searchable once the background indexer has written them.
- Requests are unchanged; a client that ignores searchIndexing behaves as before.
Memory recall can be limited to the entries about given entities.
- recall_memory and POST /memory/recall take an optional about list of 1 to 20 entity ids and return only the entries about at least one of them. With includeTenantWide, entries about no Project are returned too.
- topic is optional when about is given. Without a topic, entries are returned most recent first, with servedMode recency.
- An id you cannot see matches nothing. Requests that were valid before are unchanged.
The REST reference adds extraction over an uploaded document.
- A new operation, previewDocumentExtraction (POST /documents/{documentId}/extractions/preview), previews extraction over one section of a document's current revision. Its evidence is in the revision's coordinates.
- Extraction candidates now list possibleDuplicates: existing entities a candidate may duplicate.
- Extraction is off in the trial, so neither is callable there yet. Nothing changes for current callers.
The trial's first release.
- 15 MCP tools: search, context assembly, documents and passages, agent memory, and the graph's schema, entities and relationships.
- 28 REST operations for documents, search, context, memory and usage.