Search the memory

This journey finds a decision, a note, what the memory knows about a file, or the thread of the latest sessions. Everything here is read-only: no search writes to the memory (except adding aliases, flagged below).

What you ask the agent

You ask your agent

« Where did we talk about Git synchronization? Search the whole memory. »

→ The agent callsmem_search

• query = git sync
• limit = 10

You ask your agent

« Search for “sealed” in the knowledge notes only, archives included. »

→ The agent callsmem_search

• query = seal
• zone = 20-knowledge
• include_archived = true

You ask your agent

« Before I change this file, tell me what the memory knows about it: src/domain/session.rs. »

→ The agent callsmem_about_path

• path = src/domain/session.rs

You ask your agent

« Show me what is linked to the note 20-knowledge/runner, two levels deep. »

→ The agent callsmem_relations

• path = 20-knowledge/runner
• depth = 2

You ask your agent

« Summarize the last five sessions of kessho-docs. »

→ The agent callsmem_digest

• slug = kessho-docs
• n = 5

You ask your agent

« Add the words “it stutters on phones” to the note 20-knowledge/runner. »

→ The agent callsmem_alias_add

• path = 20-knowledge/runner
• aliases = ["it stutters on phones"]

A question about the past ("what did we decide about…?", "where did we talk about…?") is enough for the agent to search: the delivered procedures (see Connect the agent) invite it to.

What the agent does

1. Search for text. mem_search searches the whole vault and returns one match per line, with its context. The query pattern is a regular expression. The options, all optional:
   • case_insensitive: ignore case (on by default);
   • fold_accents: "reglage" finds "réglage" (on by default);
   • fold_separators: spaces, hyphens and underscores interchangeable, "deploy kit" finds deploy-kit; no effect if the pattern contains a regex metacharacter (on by default);
   • search_names: the file name counts too, those matches carry line_number 0 (on by default);
   • zone: restrict to one zone (00-inbox, 10-episodes, 20-knowledge…);
   • include_archived: include 10-episodes/archived/, excluded by default;
   • limit: 50 matches by default, 500 at most.
2. Know what the memory knows about a file. mem_about_path takes the path of a file in the code repository, relative to its root (an absolute path is refused), and returns what the vault knows: the map facts describing it, the manifest that established its technology, the framing document that concerns it. The catalog asks for it to be called before touching a file you do not know. Knowing nothing is not an error: the answer then says how many notes cite any file at all.
3. Follow a note's links. mem_relations takes a note's path (relative to the vault, with or without .md) and returns what is more general (up), more specific (down) and neighbouring (side). direction is up, down, side or all (default); depth goes from 1 to 6 (default 2). Beyond 200 notes the result is marked truncated; links whose target does not exist are named.
4. The thread of sessions. mem_digest summarizes the n most recent archives of a project (5 by default, 50 at most): date, subject, excerpt. An archived project is only summarized with from_archived: true. To read a whole archive, see mem_read_archive in Resume a project.
5. The whole graph. mem_graph_export exports the notes and their relations; the payload is large, and levels keeps only the levels asked for (domaine, concept, projet, note, archive).
6. Help find it again. mem_alias_add writes: it adds everyday-language aliases to a note (the symptom, the need) so it can be found with those words. Aliases already there are never lost.

What you see

An agent search changes nothing on screen: the result is in your conversation with the agent. The application has its own ways to find a note: Search in the sidebar (see Home) and the graph.

Only mem_alias_add writes: it lights up the freshness indicator like any agent write (see Zones).