# Your Agent Remembers You. Who Remembers the Company? > Personal memory can follow one user between tools. Institutional memory needs owners, effective versions, approvals, and evidence. Published: 2026-07-23T04:00:00Z Updated: 2026-07-24 Author: Archie Norman (Founder, MLX) Category: Twin Tags: twin, agent-memory, institutional-memory, knowledge-graphs, provenance, enterprise-ai, governance Canonical URL: https://mlx.systems/blog/who-remembers-the-company ## TL;DR - An agent remembering a user's preferences is personal continuity; it is not the same as the company retaining an approved definition. - Institutional memory must distinguish proposals from decisions, preserve historical versions, name accountable owners, and remain available to every authorised interface. - MLX provides durable user, channel, and organisation memory, but only the Twin governs company meaning through versioned products, publication evidence, and citations. - Temporal graphs and longer context can improve recall, but recall cannot decide which contested business definition has authority. --- Your agent may soon remember how you write, which projects matter to you, what you decided last week, and which explanation you rejected yesterday. That will be useful. It may also create a dangerous illusion: because the agent remembers a person, the company now remembers itself. It does not. A personal agent can preserve continuity for one user. A company needs something stricter. It needs definitions that survive staff changes, model changes, application changes, and disagreements between teams. It needs to know which statement was a proposal, which was approved, when the approval took effect, and what evidence supported an answer at the time. Personal memory is recalled. Institutional memory is committed. ## The Appeal of an Agent That Remembers Every conversation with a stateless assistant begins with a tax. You explain the project again. You repeat the relevant constraints. You remind it which style you prefer and which approach already failed. When the context window fills or the session ends, much of that continuity disappears. Agent memory systems are trying to remove that tax. Some store summaries. Some retrieve semantically related facts. Others build graphs of entities and relationships so an agent can reason about how a person's world changed. Graphiti is a useful example. It represents facts and relationships with temporal metadata rather than treating knowledge as a static pile of text. The accompanying Zep research argues that enterprise agents need to synthesise changing conversational and business data while preserving historical relationships. [Read the Zep paper](https://arxiv.org/abs/2501.13956) and [the Graphiti launch discussion](https://news.ycombinator.com/item?id=41445445). This is real progress. An agent that knows someone changed jobs should not continue answering as if the old role were current. An agent that remembers an architectural decision can avoid reopening the same argument in every coding session. But the success of personal memory makes the organisational distinction more important, not less. ## The Company Is Not a Larger Person A company does not have one coherent stream of experience. Finance, sales, delivery, and operations may each hold a defensible definition of the same term. Two definitions can be valid for different decisions. A policy can apply to one legal entity but not another. A metric can change after a board decision while historical reports must retain the old treatment. If an agent extracts a fact from each conversation and tries to reconcile them, it may produce a smooth answer by erasing the disagreement that mattered. Consider “active customer.” One sales leader may mean any account with an open opportunity. Finance may mean a customer with recognised revenue in the period. Customer success may mean an account with an active service plan. The newest conversation is not automatically the authoritative one. The most frequently repeated definition is not necessarily approved. A graph edge saying `Customer -> IS_ACTIVE` still needs a business context, an owner, and an effective period. The company does not need its agent to remember the phrase more confidently. It needs to govern the decision the phrase represents. ## Three Ways Personal Memory Fails at Company Scale The difference becomes clearer through the failure modes. ### 1. Recollection has no approval state A memory system can store that someone said, “renewal risk starts at 90 days.” It may not know whether that was an idea, a draft rule, an instruction for one account, or the policy approved for every report. Turning every confident statement into durable company context promotes conversation into authority. An organisation needs at least two states: proposed and approved. In practice, it needs more. Discovery, review, validation, publication, activation, and evaluation answer different questions and should leave different evidence. ### 2. Recollection can overwrite history Personal memory often aims to keep the latest relevant fact. If a user now prefers Puma, the old Adidas preference can become historical context. Business reporting cannot always work that way. If the margin definition changed in July, an answer produced in June must still be interpretable under the June version. A later approval should not rewrite an earlier run, event, certificate, or evaluation receipt. Institutional memory needs immutable historical identity, not only a current best guess. ### 3. Recollection belongs to the wrong custodian If a crucial business rule lives only in one employee's assistant, who owns it? The employee may leave. The agent vendor may change retention policy. The memory may be visible only in one interface. Another employee's agent may remember a conflicting version. Company knowledge cannot depend on which person's assistant happened to hear it first. It needs an organisation-scoped home and an accountable owner. ## What It Means for a Company to Remember Institutional memory is often described as documentation. Documentation matters, but a folder full of prose is not the complete answer. For an agent to rely on a business definition, the company needs a committed artefact with several properties: - **Identity:** the product has a stable name and version. - **Ownership:** an accountable person or role can approve or reject it. - **Scope:** the definition names the organisation, grain, fields, sources, and permitted use. - **Effective history:** a new version does not erase the version used before it. - **Validation:** the published result has evidence tied to the exact contract. - **Runtime binding:** an agent is constrained to the activated product rather than merely advised to prefer it. - **Citation:** a completed answer can identify which version and relation were used. That is closer to committing code than saving a chat summary. A commit does not claim the code is perfect. It establishes identity, authorship, history, and a reviewable change. Institutional memory needs the same discipline around business meaning. ## The Twin as Institutional Memory The Twin is MLX's answer to the company side of this problem. It does not try to remember every employee's conversations. It turns connected evidence and reviewed business definitions into governed products. As [the opening article in this series](/blog/agent-is-the-easy-part) explains, the Twin moves a candidate definition through durable discovery, human review, publication, activation, and optional evaluation. The important memory property is that each state has its own identity and evidence rather than becoming one mutable “current answer.” That sequence distinguishes learning from deciding. The Twin may discover that several source columns and historical queries point toward a plausible revenue product. That is useful evidence. It is not approval. The proposal becomes institutional memory only when the organisation accepts the definition and publishes a version that agents can bind to. Activation is separate again. A product can exist without being the active product an agent is allowed to use. Evaluation is separate from both, so a published definition and an agent's success with supported questions are not collapsed into one claim. Institutional memory is not one database table. It is the lifecycle that tells you what each artefact is allowed to mean. ## A Receipt Is Part of Memory Remembering a definition is only half the problem. The company also needs to remember what happened when the definition was used. The governed Twin query path can return a structured product citation derived from run events. It identifies the product, version, and published relation used by the run. Execution certificates record a separate set of runtime facts. That evidence matters when someone asks: - Why does this quarter's answer differ from the answer produced in May? - Did the model change, or did the product definition change? - Was the same relation queried? - Was the source freshness different? - Which organisation and capability boundary applied? Without product and execution identity, the organisation remembers only the final prose. It cannot distinguish a changed fact from a changed interpretation. A receipt is not a truth oracle. It does not prove that an upstream system contained the correct transaction or that every aggregate can be traced to a source row. It preserves the custody of the answer: what governed product was selected, what ran, and which evidence was observed. That is part of institutional memory because it makes past decisions inspectable rather than legendary. ## What About Agent Memory? MLX persists tasks, sessions, runs, and events. It also provides durable memory tools that can store and recall bounded facts across sessions. Those memories have explicit scopes. A user memory can preserve a person's stable preference or correction. A channel memory can preserve durable context for one workspace. An organisation memory can preserve a shared decision or context, but it is submitted for human approval before it becomes active and recallable. This is more than session continuity, but it is still not the same thing as a Twin product. A durable memory can record that someone prefers concise answers or that a team made a lasting decision. It does not publish a stable warehouse relation, product contract, validation evidence, activation record, or structured product citation. It cannot promote a remembered statement into the organisation's approved definition of revenue. That separation is deliberate and useful. The user's chosen agent may also carry personal context between products. MLX memory can preserve scoped continuity inside MLX. The Twin preserves the organisation's committed meaning. A session preserves the state of the task. These are different kinds of continuity with different owners and risks. Trying to make one memory mechanism serve all of them would blur private preference, transient work, shared context, and approved company state. ## Bigger Context Does Not Create Authority Even perfect recall would not settle the authority problem. A context window may contain an old definition, a new definition, a draft, a correction, and a project-specific exception. Remembering all five preserves the disagreement; it does not ratify one of them. Temporal knowledge graphs and governed products are therefore complementary. One preserves change and improves recall. The other determines which interpretation is published, active, and available to a particular caller. The security consequences of deciding what enters that window are the subject of [Context Is a Security Boundary, Not a Token Budget](/blog/context-is-a-security-boundary). ## The Wiki Still Matters None of this makes wikis, meeting notes, policies, or data dictionaries obsolete. They are sources of institutional knowledge. They help people explain why a definition exists and preserve context that cannot be reduced to a measure. The mistake is asking prose to enforce itself. A wiki page can say which metric is preferred. It cannot, by itself, prevent an agent from querying a different relation. It may not carry an effective version. It may contain a comment that looks as authoritative as the approved text. It may not be scoped to one organisation or reporting period. MLX ambient context can carry organisation-owned policies, dictionaries, and validation guides when they are tagged, named, scoped, and auditable. Contradictory or incomplete material should produce an inconclusive result rather than silently becoming a durable instruction. The wiki contributes evidence and explanation. The governed product supplies the binding. ## Who Should Remember What? The emerging agent ecosystem will probably contain several kinds of memory: - user-owned memory that follows a person between tasks; - session state that lets long-running work pause and resume; - application records that remain authoritative for their own transactions; - temporal knowledge systems that preserve changing relationships; - governed company products that define how shared evidence should be interpreted. The goal is not to collapse them into one universal memory. It is to give each kind the right custodian. Your agent may remember you extremely well. That can make it a better interface to your work. The company still needs to remember itself in a form that no individual agent is allowed to improvise: versioned, owned, approved, and accompanied by evidence. That is not recollection. It is institutional memory. --- **Previous: *[The Agent Is the Easy Part](/blog/agent-is-the-easy-part).*** Why company context becomes more valuable as models and agent interfaces become easier to replace. **Read first: *[You Are Not the Front Door](/blog/you-are-not-the-front-door).*** Why the user's agent and the company's interface to itself are different products. **Next: *[A2A Solves How Agents Talk. It Does Not Solve Who Is Allowed.](/blog/talk-is-not-authority).*** Why interoperability needs a separate chain of custody for authority.