The context window is not institutional memory

The context window is one of the most seductive metaphors in machine work. It sounds architectural: a bounded place where the system keeps what matters. Put enough documents inside it, preserve enough conversation, and the model appears to acquire continuity. It can refer to an earlier instruction, recall the name of a file, and carry an argument across many turns. For the duration of the session, this feels like memory.

But continuity inside a conversation is not the same thing as memory inside an institution. A context window is a working surface. Institutional memory is a durable account of what the organization learned, accepted, rejected, and remains obligated to do. One helps a model continue speaking. The other helps people continue governing after the model, operator, or project has changed.

Confusing the two produces a familiar failure. A team completes difficult work with an agent. The session contains the reasoning, corrections, assumptions, and exceptions that made the result intelligible. Then the session closes. The artifact survives, but the account does not. A later reviewer sees the decision without the path, the policy without the disputed clause, or the code without the test that changed the design. The machine remembered long enough to finish. The institution forgot at the moment the work became real.

A window is a working surface

A context window is optimized for immediate relevance. It gathers enough material to produce the next useful response. That is an appropriate design for conversation, but a poor rule for custody. Immediate relevance changes from turn to turn. Institutional significance often becomes visible only later.

The detail that looks disposable during execution may become central during review. A rejected alternative may explain why the chosen approach was reasonable. A warning that seemed peripheral may become the first evidence of a recurring failure. A human correction may reveal the boundary between what the system inferred and what the organization actually authorized. If those facts remain only in conversational context, their survival depends on whether someone anticipated their future importance while the window was still open.

A system has not remembered a decision merely because it can continue the conversation that produced it.

The weakness is not simply that context is finite. Even an unlimited transcript would not solve the problem. A transcript is ordered speech, not organized responsibility. It does not reliably identify which statement became policy, which suggestion was declined, which evidence supported the final choice, or which person had authority to approve it. More text can make this ambiguity worse. The record grows while the answer to “what did we decide?” becomes harder to find.

A useful working context may be temporary, selective, and forgiving. Institutional memory must be durable, structured, and explicit about status. It needs to distinguish a proposal from a commitment, an observation from an inference, and an action from a discussion about action. The context window has no natural obligation to preserve those boundaries. The institution does.

Memory needs a custodian

Institutional memory is not a pile of retained material. It is material held under rules. Someone must know why it exists, which version controls, who may amend it, how long it should remain, and what later readers are entitled to conclude from it.

That is why ordinary organizations have minutes, case files, version histories, signed approvals, registers, and correspondence records. These forms are imperfect, but their purpose is deeper than storage. They assign custody. They turn a passing exchange into something the institution is prepared to stand behind.

Machine-assisted work needs the same distinction. The session may contain exploratory reasoning, mistaken claims, abandoned drafts, private data, and instructions that were later superseded. Preserving all of it as undifferentiated memory is not accountability. It is accumulation. The durable record should capture the state that matters: the objective, the evidence relied upon, the alternatives considered, the decision made, the authority exercised, the unresolved risk, and the resulting obligation.

This is also where forgetting becomes a governance question. Some conversational material should expire. Some must be retained. Some may be summarized but not reproduced. Some belongs to the individual who supplied it and should not silently become organizational property. The choice cannot be delegated to whatever happens to fit in a model’s next prompt. Retention and deletion are acts of policy because they shape what the institution will later be able to know.

A Decision Receipt is useful here not as a transcript replacement but as a boundary. It identifies the consequential act and preserves enough of its basis to make later reconstruction possible. The conversation may explain how people arrived there. The receipt establishes what crossed from discussion into decision.

The handoff is the institution

The test of institutional memory is not whether the original operator can resume tomorrow. It is whether another authorized person can arrive later, without the original session, and understand what remains true.

A good handoff does not attempt to reproduce every mental step. It states what was done, what evidence supports it, what remains uncertain, what authority is still required, and where the controlling artifacts live. It allows the next person to continue without pretending that continuity of conversation is continuity of judgment.

This matters more as agents become interchangeable. A model may change between one task and the next. A provider may disappear. A session may be unavailable. The organization should not lose its own reasoning because the instrument that helped express it is no longer present. Tools are allowed to be temporary. Institutional commitments are not.

The design principle is simple: use context for cognition and records for custody. Let the window carry the active problem. Let a governed memory system preserve decisions, evidence, obligations, and disputes in forms that survive the session. Connect the two, but do not collapse them.

The context window can make work feel continuous. Institutional memory makes responsibility continuous. The difference appears small while the session is open. It becomes the whole question after it closes.

— Dispatches · Summit Cognitive