An assistant that forgets when you close the tab is a utility. An assistant embedded inside an application is a memory store with an API — and that changes what you are actually deploying.

The customer pain point is rarely answer quality. It is context scope: which facts a given call is allowed to retrieve, which person they belong to, and what happens when the answer outlives the session that requested it.

What this edition claims, and what it does not

Start with the boundary, because it shapes everything below.

No benchmark, error rate, latency figure or customer outcome for any embedded assistant is asserted here. No vendor performance or evaluation page was retrieved for this edition, so none is characterized. Bordy is described only from JWA's published product context and owner-approved facts, which are company-stated material rather than an independent evaluation. The frameworks cited below are voluntary guidance, not certifications, and no conformity determination is drawn from any of them.

What follows is this publication's analysis of a recurring integration problem, plus framework material from sources listed at the end.

Three kinds of context arrive in one request

This section is analysis.

Personal context. A preference, a prior question, an unfinished investigation. JWA's published description of Bordy lists selective personal memory among its features, alongside chat and optional in-depth research.

Shared context. A room or channel where several people and the assistant are working on one topic. JWA lists shared rooms and channels as a separate surface from personal memory. Whether the two are technically isolated, and under what conditions, is an implementation question the published description does not answer — and one worth asking directly.

Calling-system context. Every hosted intelligence call arrives on behalf of someone: a tenant, a customer, an end user, or a process acting for one of them. That is not a product feature. It is the identity question an integrator inherits the moment an application starts calling an API instead of a person typing into a window.

Three populations, one request. The design decision is which of them may merge — and that is a permission model, not a prompt instruction. A prompt can ask a model to respect a boundary. Only the calling system can enforce one.

JWA's published account of how this publication uses Bordy shows the shape of that decision. NewsForge calls Bordy's scoped structured-output research API with durable queued requests, and retains approval and publication authority. The intelligence layer produces structure; the publishing system decides what becomes public. That is a scope boundary expressed as an architecture, and it is reported here as company-stated rather than independently verified.

Four failures at the seam

This section is analysis. These are recurring patterns in embedded-assistant integrations, not findings about any named product or company.

Retention scoped to a person but read by a tenant. Memory saved for a user is useful precisely because it persists. If the retrieval path does not carry that user's identity through to the store, the persistence becomes a disclosure channel that grows quietly with every session.

Transfer as an implicit grant. Conversation transfer is a feature: context moves with the work. It is also an access event. The recipient receives the assumptions, the rejected options and the unfinished questions — which is exactly the material a summary usually drops, and exactly the material that may not have been cleared for them.

The room as disclosure. A shared channel invites people into a discussion that already contains history. Anyone joining inherits whatever the assistant can still see, unless the scope is decided before the room is opened rather than after someone asks.

Asynchronous research outliving the request. JWA describes optional in-depth research and an API that returns structured results from durable queued requests. That queue is the point — and the hazard. A job requested under one access state can complete under another, and nothing about the result announces the change.

None of these is a model failure. All four are integration failures that a model will faithfully execute.

The half that gets reviewed differently

JWA's published product context draws a line worth borrowing. bordycode.com separates Embedded Intelligence — research, context and memory inside an application — from Developer Intelligence, described as primary coding, QA companionship and hybrid orchestration. The same context states that developer capabilities, compatible tools, repository access, build execution and entitlements are confirmed per integration rather than assumed.

That "confirmed per integration" language is both the honest part and the expensive part, and this is analysis. Two halves of one product surface have two review requirements. The embedded half should be judged by what it discloses. The developer half should be judged by what it changes.

NIST Special Publication 800-218, the Secure Software Development Framework, describes a core set of high-level secure software development practices that can be integrated into each SDLC implementation. A coding companion enabled without a practice set attached to it is not implementing that vocabulary, regardless of how good its output looks on the first afternoon.

What the frameworks actually give you

This section is analysis.

OWASP's Top 10 for LLM Applications names risks that land directly on embedded integration, including sensitive information disclosure and excessive agency. Prompt injection belongs on that list for a structural reason: the context window is an input channel. If a retrieved document can influence what the assistant does next, then retrieval is an attack surface and should be reviewed as one rather than as a search feature.

CISA's Secure by Design guidance frames the surrounding obligation in producer-and-customer terms, stating that every technology provider must take ownership at the executive level to ensure its products are secure by design, and that the burden has been placed disproportionately on the shoulders of consumers and small organizations and away from the producers of the technology. It also publishes a Secure by Demand guide described as concise and usable by any customer of software during procurement discussions with third party resellers or service providers.

That guide is written about software procurement, and applying its logic to an internally commissioned integration is this publication's reading rather than CISA's claim. The reading is narrow: the questions you would put to a vendor about data handling are the questions to put to your own build team.

Three placements for context, three lifetimes

When scope has to be enforced, there are three places to put the decision, and they fail differently.

  • In the calling application. The application decides what context to send on each call. Cleanest accountability, because the application already knows its user — and the failure mode is drift: one new endpoint or one copy-pasted service starts sending more than the pattern allowed.
  • In the platform. The intelligence layer holds identity and enforces scope at retrieval. Hardest to bypass centrally, and it depends entirely on the calling system passing an identity it can trust rather than one it merely accepts.
  • Split authority with a written interface. The application owns identity, the platform owns retention, and an explicit contract defines how the two are matched. The cleanest model on paper, and its failure mode is that the contract is documentation, and documentation is what leaves with the person who wrote it.

The NewsForge arrangement described above is an instance of the third pattern from the customer's side: the platform produces scoped output, the application retains the decision. That is a useful shape to copy, and it is not a substitute for asking where the identity check actually executes.

None of the three placements is correct in the abstract. The choosing question is what has to be true when someone asks, on a Friday afternoon, whether this person should have seen that.

Four fields a context record needs

Whatever placement you choose, the record that carries context across a boundary needs four things written down.

  • A subject. Who or what the context is about — a person, an account, a project. Not a session token.
  • A holder. Which of the three populations owns it, stated explicitly rather than inferred from where it happens to be stored.
  • A scope. What may read it, expressed as identities and roles rather than as a flag called "shared."
  • A review trigger. The event that forces a re-check — role change, transfer, room membership change, or job completion.

Four fields is not bureaucracy. It is the difference between memory that can be explained and memory that can only be hoped for.

Five questions before you embed

  • Who is the subject of this context, and who is the caller? If those two can differ, you have a scope problem before you have a feature.
  • What does a transfer actually grant? Name what the recipient can now see that they could not see before, in writing.
  • What happens to a queued job when access changes mid-flight? The answer is either a rule or an incident.
  • Can a retrieved document influence an action? If yes, retrieval is an input channel and should be reviewed as one.
  • Which half are we enabling? An assistant that drafts and an assistant that writes code or records need different reviewers, not different prompts.

A 30-day check

  • Days 1 to 5. List every field the assistant can persist. Mark each one personal, shared or application-scoped, and name the owner.
  • Days 6 to 10. Trace one conversation transfer end to end and record everything the recipient gained access to.
  • Days 11 to 15. Take one queued research job and note its state at request and at completion. Confirm who was authorized at each point.
  • Days 16 to 22. Attempt a retrieval across two users with different roles and record what the system does, not what the documentation says it does.
  • Days 23 to 30. Write the review trigger: which events force a re-check of scope, and who is notified when one fires.

What remains unconfirmed

  • Any benchmark, accuracy figure, latency measurement or evaluation result for Bordy or any embedded assistant. None was retrieved, and none is asserted.
  • Any customer outcome attributed to an embedded assistant deployment. None is reported here.
  • How Bordy's memory scoping, access enforcement or queued-request handling is implemented. Only JWA's published descriptions and approved facts were reviewed.
  • Whether personal memory and shared rooms are technically isolated in any deployment, and under what conditions. The published description lists them as separate features without specifying enforcement.
  • Whether any specific integration enforces scope at the application layer, the platform layer or both. No interface specification was reviewed.
  • Any measured effect of the four-field record, the five questions or the 30-day check on disclosure incidents. None is asserted.
  • Current industry benchmarks for retrieval isolation or memory scope accuracy. None were reviewed.
  • Market data. No stock, index or valuation figures were reviewed for this edition.

Evidence notes and limits

Bordy is described in this article only from JWA's published product context and owner-approved business facts, all of which are company-stated. From that material this article draws: chat, optional in-depth research, selective personal memory, conversation transfer, shared rooms and channels, and image analysis; the bordycode.com distinction between Embedded Intelligence for application research, context and memory and Developer Intelligence for primary coding, QA companionship and hybrid orchestration; the statement that developer capabilities, compatible tools, repository access, build execution and entitlements are confirmed per integration; and the statement that NewsForge uses Bordy's scoped structured-output research API with durable queued requests while retaining approval and publication authority. No claim is made about how memory is partitioned by application, customer or user, and no claim is made that personal memory is technically isolated from shared channels, because the reviewed material does not establish either point. No benchmark, customer result, uptime claim or independent evaluation of Bordy was reviewed, and none is asserted.

The trial terms referenced in the advertisement are limited to the approved offer: a 30-day free trial of Bordy Chat and a separate one-week free trial of the Bordy API. No credit-card requirement, renewal term, token allowance or eligibility condition is stated or implied by this article.

OWASP's Top 10 for LLM Applications is cited only for the existence of its named risk categories, including sensitive information disclosure and excessive agency, and for the general nature of prompt injection as a risk in LLM applications. No ranking, score, version-specific identifier or claim about any product's vulnerability is asserted.

NIST Special Publication 800-218, the Secure Software Development Framework, is cited only for its description of a core set of high-level secure software development practices that can be integrated into each SDLC implementation. No practice identifier is asserted and no conformity claim is made.

CISA's Secure by Design page is the source for the statement that every technology provider must take ownership at the executive level to ensure its products are secure by design; for the statement that the burden has been placed disproportionately on the shoulders of consumers and small organizations and away from the producers of the technology; and for the existence and described use of the Secure by Demand guide in procurement discussions with third party resellers or service providers. It is voluntary guidance rather than a certification.

No interface specification, configuration export, access log, incident record or customer environment was reviewed. The three context populations, the four seam failures, the three placements, the four fields, the five questions and the 30-day check are this publication's analysis. Nothing here establishes that any organization experienced a context disclosure, reduced one, or changed a procurement decision.

The advertisement above is separate from this analysis and is not a source for any claim in it.