Mary Fung
noteMay 24, 2026

Retrieval is a judgment layer

Good retrieval is not just finding related notes. It is deciding what context is safe to use for the task.

Vector search is useful. It is not the whole retrieval problem.

In an agent memory system, the question is not only "what is semantically similar?" The better question is "what context should this agent be allowed to use for this task?"

That is a judgment layer.

Job

Retrieval selects working context from memory.

It should narrow the field before the model reasons. Otherwise the model inherits every ambiguity in the archive and turns it into confident prose.

Inputs

The retrieval layer needs:

The same note may be useful for one task and unsafe for another. A private scratch note might help the owner think. It should not become source material for a shared output.

Processing

Retrieval should combine semantic search with filters:

The system should also retrieve less when the task is narrow. More context is not always more intelligence. Sometimes it is just more surface area for error.

Output

The output is a context packet, not a pile of chunks.

A context packet should include the selected records, citations or source trails, exclusions when important, and a short reason each record was included.

That reason is not decoration. It lets the human reviewer see whether the system understood the task.

Review Question

The review question is: if the answer is wrong, can I inspect what context caused it?

If not, retrieval is functioning as vibes.

Failure Mode

The failure mode is relatedness masquerading as relevance.

The model finds notes that sound close, then writes an answer from them. The answer may be fluent. It may even be plausible. But if the retrieved context was stale, private, unreviewed, or wrong for the task, the system has already lost.

← back to the field