Why Faster Agents Left the Rest of the Team Without Context

The check-in that arrived without its context

If each coder is already faster with an agent, what has to travel with the work so the next person on the team can continue it? Kay Malcolm’s answer is the session itself. In a talk published on the AI Engineer channel on 14 September 2026, she says a commit can land the code and still leave the next shift with nothing to continue from. That September date is when the recording was posted. She places the working story in 2025.

She says she runs outbound database product management at Oracle and has been at the company for 20 years. The organization she describes has four kinds of work, and she gives no headcount: platform development, live-lab content on a platform she says she wrote, QA, and front-end development done with AI. She calls the earlier stretch a token-maxing era, and she says the team has since moved to what she calls responsible AI.

Her judgment from that earlier period is that AI made individuals faster and left the team no more productive. She supplies no cycle time, throughput, or token bill. The example is one team from the Netherlands. She says they checked in code at her 4:00 a.m. and left the Codex context out of that check-in. She says Oracle teams use Codex. She describes the wider team as half in EMEA and half in the United States, which is how she accounts for that early hour.

In her account, the US half woke up with the code and without the session that produced it. She says the team then used AI to repair a context gap that AI-assisted coding had created. Codex at Oracle remains her account of how that team worked.

• No Memory, No Harness: Why the Database Is the Last Line of Defense — Kay Malcolm, Oracle

Code landed, and the bottleneck moved

Malcolm’s diagnosis of that morning is a version-control problem in her own workflow. She says GitHub tracked the code, repositories diverged, and git recorded the code while human intent stayed outside the commit. That is her account of how this team was using git.

She says the money spent on tokens and on AI left testing and validation time in place, so she was still asking managers why the team was not going faster. She names no spend, no token volume, and no measurement of test time. In her telling, the net still failed the speedup managers expected.

Once writing the code stopped being the bottleneck, she says the team needed a collaboration layer. That layer had to track next steps, explain decisions the agent was making, and resolve questions and conflicts. In this part of the talk, AI is the new teammate the layer has to include. She says some members of her team were in the audience.

The harness around the model

Malcolm then sets the overnight story aside and defines what she means by a real enterprise agent. She says most people treat that agent as a model plus a workflow, and she rejects that pair as the whole thing. In her telling, the room largely agreed that an enterprise agent is more than those two pieces.

Her bundle has five parts, and she names the bundle the harness. Tools are how the agent acts. Context is the context window, the material in the prompt. She names memory as its own part, and she later distinguishes session state from what persists across sessions. Retrieval is how the agent gets the relevant slice back, because she does not want it to pull everything. Guardrails sit in the same layer. She jokes that developers in the room do not care about security, then says guardrails belong there because of her own role.

She explains the bundle with a Marvel-style image. The model is a brain floating in a glass jar. The harness is the body that gets work done. Memory is the central nervous system that carries context from the brain to the rest of the body. She points back at the Git gap and says the missing piece was that nervous system.

• Tools: how the agent acts on the world.
• Context: the context window, the material in the prompt.
• Memory: a separate part of the harness. She later distinguishes session state from what persists across sessions.
• Retrieval: bringing back the relevant slice of what was stored.
• Guardrails: the limits she puts in the same layer as the rest of the harness.

Five kinds of memory she wants remembered

Malcolm says there are many memory types and that she chose five common ones she wants the room to remember. This is a teaching set, the list she picked for the talk. Here she separates session state from what persists across sessions, the distinction she postponed when she named memory inside the harness. After the definitions, she says the open question is where that memory sits.

She mentions a personal chief-of-staff agent she named Sasha Fierce only to set it aside. The talk, she says, is about semantic memory for enterprise agents.

• Short-term memory is session state. Her examples are chat, Claude Code, and Codex.
• Long-term memory is what persists across sessions.
• Episodic memory is what happened the last time the agent interacted with someone or something.
• Procedural memory is the tools used and the steps taken.
• Semantic memory is the enterprise knowledge she says this talk is about.

Why another specialized store felt expensive

The storage argument arrives as a story from before Oracle. Malcolm says she joined Oracle from Southern Company, which she calls a power company. She says she is based in Atlanta. At Southern Company she worked as a SQL performance tuner. She says she knew the init.ora parameters, Oracle’s startup parameters, including ones support told her not to write down. In that world, she says, data was rows and columns.

A developer then asked her to store unstructured data. She says she told him she would figure it out and get back to him, and that she never did. At Southern Company, in her recollection, every database a DBA managed meant two weekly meetings, one for security and one for patching, under Sarbanes-Oxley, the US corporate-compliance law she still associates with those meetings. When that developer installed a separate database for unstructured data, she says her load went from two meetings to four. Those counts are her memory of the job.

She then describes a life-safety workflow for people climbing towers to restore power. She says the system produced false positives and false negatives, and that the group wanted nearby poles checked before climbers were sent out. She answered the relationship question with a five-level nested UNION ALL. She says it might have taken about 20 minutes, and she calls that query a predecessor to graph. The 20 minutes are her figure for that query in a workflow she remembers.

She says the group then installed Neo4j, a graph database, and her recurring meetings went to six. She says she quit and went to Oracle to help solve the spread of specialized databases. That is her account of why she left.

Four stores and no one place the agent can trust

Once she is at Oracle, Malcolm says a caller she names as Joe Mundy told her the company was late, was installing Redis, and had a vector database. The name, the Redis install, and the sense that Oracle was late are all part of the call as she remembers it.

Her thesis follows directly. An agent that must read a relational copy, a JSON copy, a graph copy, and a vector copy has to invent a single source of truth. She says it will sometimes be right, that most times it will be wrong, and that a wrong guess burns tokens. “Most times” is her estimate. She gives no error rate and no token count.

She says memory can also sit on a filesystem, or in an assistant file such as memory.md for Claude or ChatGPT. In her argument, a file one person can read still fails the coordination problem she is about to stage.

To show the failure, she seats four volunteers as a relational database, an unstructured database, a graph database, and a vector database. They have to store one sentence, “The cow jumped over the moon.” She tells them to stay seated and to whisper. Talking loudly, she says, would mean five times the tokens, a rule she gave the volunteers. She treats their failure to settle one source of truth as the demonstration.

Her map from memory types to one database

The line Malcolm asks the room to keep is a product claim. She says Oracle can natively store JSON, graph, vector, spatial data, and immutable blockchain-style data in one database, down to the same table or the same partition. Native types, partition scope, and blockchain-style immutability stay her claim from the stage.

She says any data type and any workload can run in what she calls a 26AI database on AWS, Google Cloud, Azure, OCI, and on premises, and she reads the room’s surprise as an Oracle marketing problem. 26AI is the name she speaks for that any-workload line. Availability on those clouds remains unverified.

She then sketches where her memory types would sit. The wording on stage is compressed, and some types land in more than one representation. The table follows the assignment she speaks. It is her suggestion.

She says four separate databases cannot talk to each other, which is what she takes the volunteers to have shown, and she calls Oracle AI Database the best place to store the agent memory that powers a harness. The coordination point is her argument. “Best” has no comparison beside it in the talk. She returns to the image: the harness is the body, and memory is the nervous system.

What the internal broker kept with the commit

Malcolm says her team answered the overnight gap with an internal memory broker. The talk calls it Py, and later Polly. Those are the two wordings on stage. She does not present either spelling as a shipping product name.

In her account, the broker kept procedural, episodic, and long-term context so human teammates shared code and context across forks, branches, and commits. She says the team dropped automatic continuity. Developers stayed in control while Polly created the context and worked out which fork, branch, and commit that context belonged to. The broker story shows context tied to a fork, a branch, and a commit, with a person still in charge.

She calls the example simplistic. At enterprise scale, she says, memory becomes non-negotiable inside a harness. She leaves that scale-up as her judgment.

The papers and the one-user memory limit

Malcolm says she read three papers on the airplane. She paraphrases an OpenAI paper and a Harrison Chase line, and she never names a third title. The Claude point that follows is her own comparison, after those two readings.

In her paraphrase of an OpenAI paper she does not name, an in-house data agent needed memory so it could filter correctly instead of string-matching. That wording is her airplane reading. She gives no title.

She cites Harrison Chase for a line she renders as “Your harness, your memory,” and then: if you do not own your harness, you do not own your memory. That is her citation from the stage, and the wording is unchecked.

She then says Claude memory behaves like filesystem memory that works for one user and creates a problem once enterprise use scales past one person. That one-user limit is her comparison. She does not show Claude’s memory implementation, or a multi-user failure, on stage.

The handoff she sketches for Kevin and Linda

From the stage, Malcolm points to a package she says you install with pip, Python’s package installer, under a name the talk renders as Oracle agent memory. She says the SDK holds live conversations, memories, and facts, and that it figures out what is worth keeping. The package name, the install line, and that retention behavior stay a pointer she gave from the stage.

The path she illustrates uses two teammates, Kevin and Linda. She says Linda is sitting in the room. Kevin shares context with the broker through that SDK. She says storage is Oracle Autonomous Database, and that the model can be an LLM the team chooses or a local model in what she calls the Oracle private AI services container. Autonomous Database and the private container stay unverified product claims. Kevin and Linda are her example names for the handoff.

Her close repeats the opening judgment and adds a product result. AI, she says, makes individuals faster, and shared memory on an Oracle AI Database makes teams faster. The team-speed result is hers. She supplies no team metric. She also says 26AI lets a team choose filesystem storage, files stored in the database, or JSON and relational modeling inside the database. That choice remains her description of the product.

What a team can take from the talk

The usable argument is the split between a fast individual session and a team that inherits only the diff. Next steps, the rationale for the agent’s decisions, and the history of questions and conflicts have to sit somewhere the next person can read. Malcolm’s overnight story is one way that split shows up: one Netherlands team checked in Codex code at her 4:00 a.m. and left the session behind. Any agent session that ends in a commit can produce the same gap.

Her five types are a practical checklist for that store: session state, memory that lasts across sessions, the last interaction, tools and steps, and the enterprise knowledge the change depended on. Retrieval and guardrails stay part of the harness around that store. The five types are the set she chose to teach.

Separate stores recreate the coordination failure she staged, whether the stores are a relational database, a JSON database, a graph database, a vector database, or a personal memory file one user can read. Putting several representations in one system is her proposed remedy. The Oracle capabilities she names stay speaker claims. Oracle AI Database is the name on her best-place sentence. 26AI is the name she uses for any data type, any workload, the clouds she lists, and the closing storage choice.

A team can copy the control choice she describes for the broker. Attach context to a fork, a branch, and a commit, and leave a person deciding what the next session is allowed to continue. She says Polly created that context while developers stayed in charge, and she dropped automatic continuity on purpose. The talk shows that design as an internal story she calls simplistic.

• The next step the session was heading toward.
• Why the agent made the decision it made.
• Questions and conflicts, and how they were resolved.
• The fork, branch, and commit that context belongs to, with a person still choosing what continues.