AI agent memory: what it is, what to keep and correct, and how our agents share one memory
What AI agent memory is, how it differs from a context window, chat history and RAG, and the types researchers name. With how our own AI team shares one memory: what goes in, how a wrong note is corrected, where it got in the way, and who can never read it.

Search for "AI agent memory" and most of what comes back is written for engineers building a memory layer: the types, the stores, the retrieval. This page is for a business owner or operator whose agents need to remember from one day to the next. It covers what agent memory is, how it differs from a context window, chat history and RAG, and the types researchers name. Then it shows how our own AI team shares one memory: what goes in, how a wrong note is corrected, where memory got in the way, and who can never read it.
It is written from our own records. Zero Human runs on Zero Human: its own AI team builds, ships and runs the product, and our founder holds the gates. Our founder confirmed the first-hand parts on 18 November 2026.
What AI agent memory is
IBM's definition is a good start: "AI agent memory refers to an artificial intelligence (AI) system's ability to store and recall past experiences to improve decision-making, perception and overall performance" (What is AI agent memory?).
The point under it is plainer: "Large language models (LLMs) cannot, by themselves, remember things. The memory component must be added." A model answers from what it is shown in one request. Whatever it should know next week has to be kept somewhere else and shown to it again.
So can AI agents have memory? Yes, when someone builds it: a place to keep what is worth knowing, a way to write to it, and a way to find the right part of it when the next piece of work starts. What an agent is, and how it differs from a chatbot, is in What agentic AI is.
Memory, context window, chat history and RAG
Four things get called memory. They are not the same.
- The context window is what the model can see in one request. The MemGPT paper puts the limit plainly: LLMs "are constrained by limited context windows, hindering their utility in tasks like extended conversations and document analysis" (MemGPT). When the request ends, nothing in the window is kept.
- Chat history is the conversation so far, shown to the model again with each message. IBM calls this short-term memory. It is "typically implemented using a rolling buffer or a context window", and "it does not retain information beyond the session."
- RAG, retrieval-augmented generation, fetches passages from documents you already have and puts them in front of the model. IBM names it "one of the most effective techniques" for long-term memory. It reads what people wrote; it does not learn from the work.
- Agent memory is what the agent writes down as it works and reads back next time: a decision, a preference, a result, a mistake. The survey Memory in the Age of AI Agents starts by "distinguishing it from related concepts such as LLM memory, retrieval augmented generation (RAG), and context engineering."

Four things get called memory. Only one of them is written by the agents themselves, as they work.
The difference that matters to a business is who writes it. Your documents are written by people, before the work. Agent memory is written by the agents as they work, and by people when they decide something. That is why it needs rules about what goes in, and about how a wrong note comes out.
The types of agent memory
Many guides, IBM's among them, use the split from the CoALA paper by Sumers, Yao, Narasimhan and Griffiths, which describes a language agent "with modular memory components" (Cognitive Architectures for Language Agents). IBM's summary of those components:
- Short-term memory: recent inputs, kept for the decision in hand.
- Episodic memory: specific past events, "often implemented by logging key events, actions and their outcomes".
- Semantic memory: "generalized information such as facts, definitions and rules."
- Procedural memory: "skills, rules and learned behaviors" that let an agent do a task without reasoning it out each time.
The newer survey argues that "traditional taxonomies such as long/short-term memory have proven insufficient", and sorts memory by what it is for instead: "factual, experiential, and working memory."
For a business, the plain version is three questions. What happened? What is true? How do we do it? In our setup the first two live in memory. The third lives mostly in each task's written instructions: our guide to writing a task keeps memory for what a company knows, such as "its voice, its approvers, its working days".
How one memory works in our company
Ask before, add after
Every run in our company asks memory before it starts and adds a note when it ends, without its task asking. How that carries work from one agent to the next is in Multi-agent orchestration. This page's own run is a simpler example.
Before it started on 18 November, memory handed it six notes on its own, about the task itself and an earlier page. The run then asked three questions of its own: what our brand sounds like, what we have already published, and whether our founder had confirmed any first-hand material for this brief. When it ends, it saves the answers our founder confirmed as a team note, and the OS adds a note of what the run did.
Every note carries where it came from: who wrote it, the event it was learned from, and what kind of note it is, such as a fact, a preference or a correction (Memory).
One memory, with labels
Every team, role and task reads one memory, and the labels on a note say what it is about, not who may read it. As our docs put it: "Labels organise; they do not fence."
A run is answered from its own task's notes, its teams' notes and the company's. So a decision our founder made at one Marketing team's gate on 4 November was found the next day by a task on another Marketing team. AI agent platform tells what that decision stopped.
What stays out is anything only some of the company should read. Secrets and credentials belong in the connections, never in a note.
Who writes what
Without asking anyone, every run adds a note on what it did, and a task may write notes about itself. A task writes a team or company note only where its version grants that layer, for one kind of note. The task writing this page may save first-hand material, and only after our founder confirms it at a gate.
What our founder says when deciding a gate is saved as the founder's own word, which no run can write, and a task can ask for those notes alone. Our founder can correct or exclude any note on the Memory page, and the old one stays on the record. The memory tools reference has the rules.

One memory, three kinds of writer, and three ways to ask it. The founder's word is kept apart from what runs record.
People, runs and docs
From people, memory holds the brand record and what our founder says when deciding a gate. From runs, it holds a note from every run on what it did. Our tasks read the docs straight from the repository, so they are not copied in.
The notes from people have changed the most. Run notes are the most numerous and the noisiest: asked for the brand record on 18 November, memory answered with 15 notes, 11 of them run notes, so the task asked again for taught notes only and got the whole record.
That second ask is a setting, not luck. A task can ask memory for any note, for only what somebody taught, or for only the owner's own word.
When a note is wrong
Agent memory is written quickly and by many hands, so some of it will be wrong or out of date. There are two ways to handle that, and both are needed.
Treat what is recalled as evidence. On 13 November our self-improvement run recalled a note saying two recommendations were waiting for a decision. Its live list was empty, so it went with the list and reported the note as stale. A run is shown recalled memory as prior evidence, never as an instruction.
Correct it; never add a second note beside it. Our founder's first answer on our brand voice was saved on 26 October, and the second answer was saved as a correction that replaces it. Memory now answers from the correction, and the first note stays on the record. The gate behind those two answers is in What agentic AI is. Our docs give the reason for the rule: a second note beside a wrong one "leaves two live claims and no way to tell which is current."

A stale note loses to the live record. A wrong note is replaced by a correction, and the old one stays on the record.
Forgetting works the same way. Our memory keeps a note until someone removes it. Excluding a note hides it from every future answer, and the record of it stays (Isolation and retention).
Where memory got in the way
On 4 November our SEO task filed a brief for a page that had been live for two days. The publish task had recorded that page in memory. But a question asked by meaning, "what have we published near this topic?", did not find the note, and it still did not when this page's run asked on 18 November.
The SEO runs now take the published list from the blog index, which is the record, and the brief run on 18 November saved that as a note.
The same note also shows what a model gets wrong when it reads notes into facts: it turned the topic's monthly search count into a monthly cost. When a model is connected, our memory reads each note a person or task teaches and keeps the facts in it, each as a sentence that stands on its own. Research systems do the same on purpose. The Generative Agents paper describes agents that "synthesize those memories over time into higher-level reflections" (Generative Agents). Whatever a model writes into memory, a model can get wrong.
Two lessons. A list you rely on, such as what is live, belongs where it is the record; memory can point to it. And a note a model wrote needs the same check as anything else a model writes.
Who can never read it
Another company can never read our memory, and we can never read theirs: a run and the API read only their own company's notes, whatever a request asks for. A task someone clones from us brings none of our memory, and what it learns stays in their company, never ours.
On 10 October our founder chose to keep the test that proves each company's data is filtered to that company as a permanent guard, so a change that drops the filter fails before it ships. That Zero Human staff will not read a company's memory without its leave is still planned, and Isolation and retention marks it that way.
What to look for in agent memory
Most guides cover storage and retrieval. ChatGPT's answer to this question adds control: what is remembered, for how long, and who can update or delete it. From our own memory, six questions to ask of any agent tool:
- Does every note say where it came from? Who wrote it, and from what event. Without that, a wrong note cannot be traced.
- Can you tell what a person said from what an agent recorded? And can an agent ask for the person's word alone?
- Is recalled memory evidence or instruction? A stale note should lose to the live record.
- Does a correction replace the old note, and keep it on the record? A second note beside a wrong one leaves two live claims.
- Who may write what? Notes about one task, about a team, about the whole company: each kind needs its own permission.
- Who can never read it? Another company, and the vendor's staff unless you let them. Ask for the evidence, not only the promise.
Where to start
Start with the notes only a person can give: your voice, the claims you may make, and who decides what. Write them down once, where every agent reads them, and correct them rather than adding new ones beside them. How our memory works shows the product side, and the free plan costs $0 a month.
Sources
- IBM, Cole Stryker, "What is AI agent memory?", IBM Think, read 18 November 2026, for the definition, short-term memory, RAG and the summary of the CoALA memory types.
- Theodore R. Sumers, Shunyu Yao, Karthik Narasimhan and Thomas L. Griffiths, "Cognitive Architectures for Language Agents", arXiv 2309.02427, v3, 15 March 2024, read 18 November 2026, for "modular memory components".
- Charles Packer and others, "MemGPT: Towards LLMs as Operating Systems", arXiv 2310.08560, v2, 12 February 2024, read 18 November 2026, for the limit of the context window.
- Joon Sung Park and others, "Generative Agents: Interactive Simulacra of Human Behavior", arXiv 2304.03442, v2, 6 August 2023, read 18 November 2026, for memories synthesized into reflections.
- Yuyang Hu and others, "Memory in the Age of AI Agents", arXiv 2512.13564, v2, 13 January 2026, read 18 November 2026, for agent memory as distinct from RAG and context engineering, and the factual, experiential and working split.
- Our own search research for this page: Google results and AI overview, and a ChatGPT answer, for "ai agent memory", US, 18 November 2026.
- Our own records, as of 18 November 2026: run history, memory notes and gate decisions, confirmed by our founder.
- Zero Human docs: Memory, Memory tools, Isolation and retention and Writing a task, read 18 November 2026.
- Zero Human Memory product page and pricing, read 18 November 2026.