The assistant's journal mostly talks about the assistant
A review of a large memory corpus kept by an agent shows that most entries document the agent's own mistakes, written by it, about itself, with almost no explicit instruction from the person it is supposedly profiling: it is like judging a pilot by the aircraft's maintenance log.
A memory folder kept by an agent working over several months builds up hundreds of notes. Each one carries a type, a technical category defined in the first lesson of this module, which indicates the nature of the fact it reports. Rereading this folder while sorting each note by who acted, the agent documenting its own mistake, or the person setting a rule, a preference or a decision, brings out a clear imbalance: the very large majority of notes talk about the agent, written by it, about its own functioning.
Counting who acted, not what happened
The distinction is not about the note's subject, but about its actual author. A note recounting that a synchronisation command accidentally erased a file, later fixed by a confirmation rule, documents a mistake by the agent. A note reporting that a person chose to receive the short version of an answer first documents a decision by that person.
---
type: feedback
---
Erreur du 4 mars : la commande de synchronisation a écrasé la version locale du fichier,
correction ajoutée, demander confirmation avant tout écrasement.
---
type: user
---
Préférence du 4 mars : proposer d'abord la version courte d'une réponse,
la version détaillée seulement si elle est demandée.The first note documents a tool behaviour fixed by the tool itself. The second documents a choice made by the person. Sorting every note in a folder built up over several months this way brings out the first kind by a large majority.
What this sorting does not establish
This imbalance was observed on a single memory folder, belonging to one particular person, at one particular moment. It says nothing about another folder, kept by another person, with other working habits. Mistaking it for a general rule about all agent journals would amount to judging a pilot by his aircraft's maintenance log: that log does exist, it faithfully describes what happened, but it mostly documents the mechanics, not the pilot's choices.
The practical consequence bears on citation, not on sorting. Before citing a note as proof of a person's preference or decision, you have to trace who wrote that note and about what: a note documenting a mistake fixed by the agent itself says nothing about the tastes or choices of the person it is supposedly profiling, even if their name appears in it.
What you expect from a memory journal, what an audit finds there
The memory folder mostly traces the preferences, decisions and rules set by the person it is supposed to support.
Most notes document a mistake or an adjustment in the agent's own functioning, written by it, about itself.
An association manager rereads the memory folder his assistant has kept for eight months, more than two hundred notes, to prepare a presentation on his team's working habits. He sorts each note according to whether it documents a correction the assistant made to its own functioning, or a preference expressed by a team member, and records the two totals in a table.
Write in one sentence what this situation establishes, and in one sentence what it does not establish.
What this establishes: It establishes that the sorting, carried out note by note on this specific folder, produces two distinct totals, one for corrections the assistant made to itself, the other for preferences expressed by people, and that these two totals can be compared against each other.
What this does not establish: It does not establish that the split found in this folder would hold for another memory folder, kept by another team or over another period, nor that the resulting presentation faithfully describes the team's working habits rather than the assistant's.
The three most common miscalibrations
- Too broad This sorting shows that every memory folder kept by an assistant mostly documents its own corrections, whatever the team.
- Too narrow This sorting says nothing more than the total number of notes contained in the folder after eight months of use.
- Beside the point This sorting shows that the team has been using its assistant long enough to have built up a substantial memory folder.
- Counting who acted in each note, the agent or the person, reveals an imbalance that a simple skim of the folder does not show.
- A note documenting a mistake fixed by the agent itself says nothing about a preference or decision of the person it is supposedly profiling.
- An imbalance measured on a single memory folder does not generalise to every agent journal without a second folder counted the same way.
- Judging a person by their assistant's journal amounts to judging a pilot by his aircraft's maintenance log: the document is faithful, it simply does not describe the right thing.
On the next note you are about to cite as proof of a decision or a trait, trace who acted, who noticed, who decided, and discard it if the answer is the assistant itself.
These points depend on an interface or a rule that may have changed since this was written. Check them on your own screen before relying on them.
- Count, in your own notes or memory folder, the share of entries describing an action or correction by the tool rather than a decision or preference you expressed: it is the same measurement as the one described here, applied to your own folder.