The ingestion box in an agent memory system should be boring on purpose.
It should not decide what is true. It should not rewrite the record. It should not turn a half-thought into a durable belief because the embedding model found it semantically convenient.
Its job is capture plus structure.
Job
Preserve the raw input and make it easier to review later.
That distinction matters. A useful personal system needs the original source, the extracted structure, and the reviewed memory to remain separate. If those collapse into one object, the system starts treating saved fragments as settled knowledge.
Inputs
Raw intake can be a note, link, file, meeting fragment, transcript, screenshot description, pasted message, voice memo, draft, decision, task, or stray idea.
The system should assume the input is untrusted until reviewed.
Processing
The intake step should create a raw record with:
- source type
- capture date
- original location
- author or origin
- project or domain guess
- people, systems, and topics mentioned
- sensitivity marker
- short abstract
- candidate decisions
- candidate tasks
- candidate reusable facts
- confidence marked as unreviewed
The important word is candidate. The system can suggest structure. It should not promote structure into truth.
For example, if a meeting note says "maybe we should use a shared agent," the intake layer can tag it as an idea about orchestration. It should not store "we decided to use a shared agent."
Output
The output is a raw intake record.
It is searchable enough to be found. It is structured enough to be reviewed. It is not yet durable memory.
Review Question
The review question is not "is this related to something?" Everything is related to something.
The question is: should this be reused later by an agent?
If yes, it needs promotion. If no, it can stay as archive.
Failure Mode
The common failure is letting the inbox become memory.
That feels efficient at first. The system has more context. It answers more often. It appears to know you.
Then it starts retrieving old guesses, private scratch notes, stale project states, and abandoned ideas as if they were current operating context. The problem is not that retrieval failed. Retrieval did exactly what it was asked to do.
The architecture failed before retrieval ever ran.