Connecting your applications does not create memory

Everything is there, but the story is missing

Imagine a work meeting. The appointment is in the calendar, the preparation in email, the guest in a contact record, the presentation in shared storage and the agreed actions in a task manager. Depending on its permissions, an AI assistant can query several of these sources.

All the traces exist. Three years later, a new manager asks: “Why did we ultimately choose this option?” She finds the presentation, date and actions. The reason for the choice remains missing.

Perhaps a constraint was discussed aloud. Perhaps an alternative seemed too risky at the time. These are hypotheses, not established explanations. Each application did its job; none of the traces necessarily preserved the collective reasoning.

Retrieving is not remembering

Three levels help name the gap. A trace is a preserved item: a message, note, document or meeting record. It establishes that content was recorded, not that every claim is accurate. Access is the ability to retrieve such items through search, connectors or an assistant. Memory organises traces with their origins, participants, dates and relationships so their meaning remains understandable and open to revision.

Retrieval locates information. Aggregation brings information together. Building memory also preserves what explains its connections. A folder containing every meeting file can still fail to explain why the third proposal was rejected.

This extends an older question. In 1991, James P. Walsh and Gerardo Rivera Ungson examined organisational memory through information acquisition, retention and retrieval [1]. Our interpretation applies that question to an environment of easier access: having traces does not remove the need to organise understanding.

What documents do not always say

Ikujiro Nonaka describes knowledge creation as a dialogue between tacit and explicit knowledge [2]. A document does not necessarily contain everything that enabled action: practice, experience and judgement may remain with people.

In our meeting, the minutes say “option B selected”. A participant might explain that option A depended on a partner who was unavailable. Collected today, this explanation would be dated, attributed testimony. It should not be presented as part of the original minutes.

Preserving context can begin with a simple gesture: ask which constraint made the difference when the decision is made. Accept that some reasoning may remain incomplete rather than rewriting it afterwards.

AI changes access, not the nature of the problem

Retrieval-augmented generation, or RAG, combines retrieved documents with a system that generates a response. Patrick Lewis and his co-authors demonstrate improvements on the tasks evaluated [4]. This can help answer questions from a corpus; it does not certify every proposed relationship.

It would nevertheless be inaccurate to claim that current systems cannot reconstruct complex context. GraphRAG, studied by Darren Edge and his co-authors, extracts a graph and summarises groups of information. On the global sensemaking questions evaluated, the authors report improved comprehensiveness and diversity compared with conventional RAG [5].

A precise distinction follows. Reconstructing context from documented relationships can be useful. Inferring a decision’s motive from converging clues remains an inference. If the sources cannot establish a conclusion, fluent wording should not hide that limitation.

An assistant might write: “The exchanges suggest a scheduling constraint, but no retrieved document confirms that it determined the choice.” This leaves a question open. It preserves memory better than a confident explanation with an invented motive.

What memory needs to preserve

Provenance describes where information comes from and how it was produced. The W3C’s PROV framework relates entities, activities and the people or agents involved [3]. For our meeting, that includes identifying the minutes’ author, their date and the documents behind a summary.

Context identifies the constraints and information available when the choice was made. Relationships connect documents to decisions and participants. Being listed as an email recipient does not, however, prove that someone read it.

Time distinguishes what carried authority then from what has changed since. Status separates documented facts, testimony, interpretation and hypotheses. Alternatives and the reasons for rejecting them preserve reasoning beyond the outcome.

Keeping everything does not answer these questions. Useful memory requires selecting what deserves transmission, defining access and correction rights and allowing deletion. Essential connections should also remain understandable with exported content when tools change.

Put it into practice: the five-year test

Choose an important decision made in the past six months. Imagine someone outside the project having to understand it in five years without calling you. For each point below, identify a source or write “missing”.

1. The final decision: what was to be done, and within what scope?

2. The people involved: who proposed, contributed and decided?

3. The supporting documents: which versions were available at the time?

4. The main alternatives: what other possibilities were actually considered?

5. The reasons for rejecting them: which constraints or arguments determined the choice?

6. The status of the explanations: what is documented, testified, interpreted or still uncertain?

If the first four are easy to retrieve but the last two are not, you may have many traces without a truly transmissible memory. This is an editorial exercise proposed by Mnemory, not a scientifically validated measurement instrument.

To act, complete a short decision note: choice, reasons, alternatives, sources, date and responsible person. If written afterwards, say so and attribute the recollections collected. Give the file to a colleague outside the project. Their questions reveal connections that still need documenting.

Preserving what connects the traces

The next digital challenge may be less about connecting more applications and more about preserving what connects their traces and gives them meaning over time.

This difference between accumulated traces and contextual continuity is precisely what we seek to explore with Mnemory: a framework for reflection and experimentation centred on sources, relationships and status, without confusing that ambition with a guarantee of exhaustive memory.

Next time a decision is recorded, let us add a question: will someone still be able to understand why?