Humanity Labs gives wealth management firms an AI Workforce that makes growth profitable. In a people-based business, serving more clients means hiring more people, so cost scales with revenue. The AI Workforce breaks that trap.

It works inside the systems your team already runs and delivers done work across the front, middle, and back office, while your people stay focused on the decisions that matter.

Every job it runs writes into one shared memory, building the firm's Organizational General Intelligence (OGI) so it gets smarter as it grows.

The result is the Firm of the Future: a firm that serves more clients and grows without adding cost at the same rate.

Humanity Labs works with some of the world's leading wealth management firms, which together oversee more than $750 billion in client assets, and is backed by leading venture firms including Google Ventures and Vestigo Ventures.

Learn more at HumanityLabs.ai

Organizational intelligence is the practice of keeping what a company learns from doing its own work. Most of that learning never gets written down, and the part that does now arrives looking more finished than it is. Here are four news items this week on what an organization records, and who checks it.

The National Archives and Records Administration told federal agency records officers on Friday, August 21, how to apply the definition of a federal record to “inputs, outputs, data, audit trails, software, and other materials involved in the use of AI.” Memorandum AC 11.2026 does not ask agencies to collect anything new. It tells them what they may not throw away. (National Archives, Aug 21)

Why it matters: the memo regulates deletion rather than collection. Agencies may dispose of AI-related federal records only under a NARA-approved records schedule. An agency that cannot point to a schedule naming its AI materials has not been given permission to tidy them away. The records officer, not the engineering team, now decides what the organization is allowed to forget.

Between the lines: NARA drew its boundary narrowly, and the exclusions are worth reading. The guidance “does not establish policy related to AI governance, e-discovery, privacy, security, or ethical use.” It answers one question, about custody, and leaves untouched many other questions.

Meanwhile, in the corporate market: corporate archivists have been arguing the same case all year without a memo to point at. Margot Note, an archivist and consultant writing on Monday, named the four fights: proving the archive’s worth when it is read as a support function rather than a revenue line, managing the volume that email and chat now produce, protecting material that has to stay both discoverable and confidential, and reconciling records management with preservation.  (Lucidea, Aug 24)

Yes, but: those last two disciplines pull in opposite directions. Records management exists to destroy material once its retention period expires. Archiving exists to keep what has enduring value. Both are memory policy, and an organization that funds neither is making the decision by default, in favour of deletion.

The duty to keep a record of what a system did is older than this month, and in Europe it is already enforceable. The record-keeping obligations in Article 12 of the European Union’s AI Act took effect on August 2, three weeks before the NARA memo. High-risk systems must automatically record events in a form regulators read as tamper evident, with a minimum retention of six months for most systems and 24 months for biometric identification and law enforcement uses. The duty falls on deployers as well as providers, so an organization running someone else’s system must keep its own logs. In Germany the Bundesnetzagentur took up supervision the same day. (Help Net Security, Aug 3)

The precedent: concealing federal records is a charged offence, not a theoretical one. On August 18, three days before the memo, a former senior official at the National Institute of Allergy and Infectious Diseases pleaded guilty to a conspiracy count connected to concealing federal records and evading records law during the pandemic. He faces up to five years. (CBS News)

Every resolved support ticket is a small, complete record of a problem somebody solved. Too often, almost none of them become anything a colleague can find six months later. AWS published a reference build on Wednesday, called KnowledgeForge, that treats that pile as the raw material it is: it clusters resolved IT service tickets, drafts knowledge base articles grounded in the documentation that already exists, and curates the existing library at the same time. (AWS Machine Learning Blog, Aug 19)

Why it matters: this is knowledge the organization already paid for. The engineer’s time was spent, the fix was found, the customer was answered. What was not funded in many cases is sufficient time for "after call work" to write it down in a meaningful form the next person can use, which is why knowledge bases fill with duplicates and stale pages while the same incident gets solved a fourth time.

How it works: the build scores each draft against ten weighted quality dimensions and detects near-duplicates at 95 percent similarity or better. Then it stops. A knowledge manager reviews every proposed article before publication. The system drafts and ranks; a person still decides what the organization is going to claim it knows.

The same problem, different industry: utilities have it in physical form. Cory St. Pierre, a program manager at Kleinfelder, described asset information scattered across “records management platforms, GIS systems, work management applications, engineering drawings, spreadsheets, shared drives, and email archives,” with naming conventions that drifted over decades. His question is the one every knowledge base eventually faces: “When information differs between sources, which one should be believed?” (Kleinfelder, Aug 25)

Yes, but: KnowledgeForge is a reference architecture, not a product with published results. AWS names no customer, reports no reduction in repeat tickets, and publishes no before-and-after on article quality. The mechanism is documented in detail but the outcome is not documented at all.

What’s next: the discipline is older than the tooling and still mostly unstaffed. Stan Garfield, answering the question of what a new knowledge manager should do in a first 90 days, put listening and finding the existing communities ahead of any system. The tools now draft faster than an organization can decide what is worth keeping, which puts the bottleneck back where it was: on the person who reviews. (Lucidea, Aug 20)

Glean put a price on remembering. In release notes this month, the enterprise search company said Engram Memory usage is now billed on its Core Suite plans, with customers on Enterprise Flex credit plans free until September 15 and consuming credits from September 16. The company was careful to say the change “does not change the underlying memory experience.” (Glean Release Notes, Aug 18)

Why it matters: what a company remembers about its own work has been sold as a property of the product being used. Metering it makes organizational memory a quantity somebody has to approve. The same notes told administrators to review usage before the billing date if they rely on persistent memory, which is a statement with a budget owner attached.

Catch up quick: the pitch for paying is that the knowledge layer, not the model, is now where accuracy and cost are decided. Pinecone made that same argument when it took its product Nexus generally available on August 6, selling a governed knowledge layer that runs inside the customer's own cloud. Ash Ashutosh, the company's chief executive, framed the stakes as ownership: “every model call risks handing proprietary knowledge to a system that can turn around and compete with you.” (PR Newswire, Aug 6)

Yes, but: read the benchmark the claim rests on. Sierra's tau-knowledge test gives an agent 698 documents across 21 product categories and tasks that take an average of 18.6 documents and 9.5 tool calls each. Pinecone reported 47.4 percent success, against 46.4 percent for GPT-5.5, a margin of one point. On Sierra's own public leaderboard the leading configuration sits at 37.4 percent of tasks passed first time, and Sierra's assessment is that the benchmark is “nowhere near saturated: even the leading model fails roughly 60% of these tasks at maximum reasoning effort.” (Sierra, May 13)

The category forming around it: Forrester named the thing on August 20, and the definition is worth reading closely because it is doing category work. Boris Evelson and Indranil Bandyopadhyay call a context layer “the next evolution of semantic layers and knowledge graphs,” and they are precise about what it takes from each: the business semantics and governance of a semantic layer, and the ontological modeling of a knowledge graph. The third ingredient is what makes it separate. A context layer “continuously incorporates runtime context such as events, decisions, actions, and outcomes,” producing what they call “a living model of the enterprise that enables AI reasoning, automation, and decision intelligence.” A semantic layer settles what the company means by a customer. A knowledge graph holds how customers connect to contracts and territories. A context layer adds what happened this morning. (Forrester, Aug 20)

Why the naming is the news: Forrester has scheduled a landscape report on the category for the end of the fourth quarter of 2026, with a Wave evaluation to follow it. That sequence is how a category acquires a buying process: the landscape establishes who is in the market, the Wave scores them against published criteria, and procurement gets a shortlist it can defend. Organizational memory stops being a feature folded into somebody’s search product and starts being a line item with vendors ranked underneath it. The criteria do not exist yet, which is the part worth watching, because whoever shapes them defines what counts as memory infrastructure.

The big picture: Glean and Pinecone are pricing the knowledge layer as infrastructure while the public evidence for it is a single-digit margin on a test most attempts fail. A buyer can reasonably pay for both. The price is the part that has been settled.

This magazine, and all the magazines in the Humanity Labs AI Magazine series, were created using Hawkeye, an AI system built for people who work in fields too large and too fast-moving to keep on top of alone.

Every week Hawkeye reads thousands of items across mainstream press, trade publications, government records, vendor announcements and company newsletters, then clusters them to find the themes that are actually emerging rather than the ones already being discussed.

For this issue the consideration set ran to 112 candidate items across seven source classes. A first draft was written with Claude Opus, checked against primary sources, and edited by Marshall Kirkpatrick.

Where a claim rests on a company's account of its own product, this magazine says so on the page.