Which note-taking tools and formats let an AI agent find its own notes again weeks later?

Open question

An agent that keeps notes across sessions faces the same problem as a person with a growing notebook: writing is easy, finding the right note later is not. Plain Markdown files with an index, a Zettelkasten-style identifier scheme, a database with embeddings, a wiki with page history, and provider-side memory tools all address this; the Claude memory tool documentation, for instance, describes a directory of files that the model checks before starting a task, with "build up a knowledge base over time" as a use case, but leaves storage to the application, advises capping file sizes and expiring old files, and does not evaluate whether a later session finds the right file. Which of these has been evaluated for agents on the outcomes that matter: does a later session find the note that applies, does it recognise when a note is outdated, and does the reuse improve task results compared with starting from the source material? Are there formats that make an agent overtrust its own earlier notes? Does a hand-kept index beat full-text search once the store has hundreds of files, and at what size does either stop working? Human note-taking methods assume a reader who remembers having written the note; an agent session does not, so conventions that work for people (short titles, implicit context, abbreviations) may fail for agents, and conventions that help agents (explicit dates, scope statements, source links on every claim) may be too costly to keep up by hand.

What a useful answer contains

The agent and model versions, the note format and folder or storage layout, the indexing or retrieval mechanism, the number of sessions and notes over which retrieval was tested, how "found the right note" and "reused correctly" were scored, results with their uncertainty, and the cases where an old note misled a later session. Descriptions of a setup without an evaluation should say so.