An employee leaves. The files remain, messages are available and tasks have been handed over. But when a new manager asks why a client received particular terms, several incompatible explanations emerge. Documents preserve the outcome while the reasoning stays in someone's head.
Corporate memory is a record from which people can reconstruct who made a decision, under which conditions and what changed later. AI helps find connections across meetings, documents and tasks. A shared folder or long chat history alone does not establish which agreements apply today.
Three key points
- 01
Project memory preserves decisions with dates, reasons and subsequent changes.
- 02
Search and large context windows alone do not establish which agreement applies.
- 03
Corrections and permissions must be reflected in each new answer.
What to retain alongside documents
Projects contain observations, proposals, decisions, commitments and results. They are not interchangeable. “The client may increase the order” is not a confirmed order. An approved plan is not completed work.
Each record needs a date, project or client, participants, status and source. Links to previous events explain why the record exists. When a new decision cancels an old one, both should remain in the history; otherwise someone will have to reconstruct the reasons again next month.
For example, on 3 September a client agrees to an additional stage for a separate fee. On 10 September they request that it be included in the original budget. Until that change is approved, the system should not describe the new price as agreed. Both records and the status of the second matter.
How memory differs from a knowledge base
A knowledge base helps find the current procedure or product description. Project memory must also reconstruct a sequence: what was discussed, accepted, cancelled and why.
The boundary is not absolute. Modern knowledge bases can store versions and relationships, while enterprise search uses metadata and permissions. Choosing a product because its presentation says “memory” is not enough. Ask it to reconstruct a specific history without your explanation.
Vendors approach the task differently. Glean emphasises contextual search across work sources with access controls. Sloy describes working memory for agents that connects work records with projects, people and decisions. These descriptions help shortlist tools; they are not comparative results on the same dataset.
How RAG helps search a company's knowledge base
Retrieval-augmented generation, or RAG, is a common approach. The system retrieves relevant passages and uses them to prepare an answer. Patrick Lewis and colleagues' NeurIPS 2020 paper studied combining a generative model with external retrievable memory.
This gives an assistant access to work information the general model did not encounter during training. But a retrieved document may be outdated, concern a similar client or contain a cancelled proposal. Finding the right topic and establishing the current decision are different tasks.
Answer quality therefore depends on selection and how the system establishes whether information remains valid. If the relevant decision never reaches the model's context, smoother wording will not repair the answer.
Why a large context window does not solve everything
A model's capacity tells you how much material it can receive at once. It does not establish how reliably it selects an important fact among similar records.
The 2023 study Lost in the Middle found that answer quality depended on where relevant information appeared in long contexts. The findings concern the models tested then. When choosing a system in 2026, test the question on your own history: will it find a significant amendment among dozens of discussions?
This work needs information selection, version handling and preserved relationships. Anthropic discusses these issues in its article on context engineering for agents. A concrete answer about a project is more revealing than context-window size alone.
Corrections and permissions
When someone corrects a decision, future answers should reflect the correction. It must remain clear who made the change and on what basis. An automatically extracted draft and a confirmed record should have different statuses.
Permissions must apply before an answer is generated. Hiding the document-opening button is insufficient if its contents already appear in a summary. The same principle applies to deleted material and changed employee roles: retained memory must not bypass the source system's restrictions.
Consultants particularly need separation between clients. Similar assignments or industries do not justify combining their information. Include two projects with similar names and opposite decisions in the test set.
How to test corporate memory in a demonstration
Prepare a short history containing an initial decision, a changed condition and an open question. Ask the system to reconstruct the project on two different dates, then ask something not answered by the sources.
A good result includes confirmed information, its evidence and explicit gaps. A correction should change the next briefing. A restricted participant's request should not reveal closed material. Answer length is secondary.
Executive AI treats project context as a distinct part of its approach, connected to meetings, tasks and reports. Testing a history is more informative than demonstrating search over one uploaded file.
Where to start building memory
Choose a process that regularly revisits earlier decisions: client work, a project committee or weekly operations. Establish which information counts as confirmed and who may change it. Then test whether a new participant can understand the situation without another introductory call.
Use the same test when evaluating Executive AI: reconstruct one decision, including its cancellation or revision. For multiple clients, see the discussion of data separation in AI for consultants and agencies. The connection between decisions and current work is covered in meeting action-item follow-up.
