Skip to content
Jensen

Memory and knowledge search

Project memory stops each session rediscovering old ground. Knowledge search finds it again.

What a memory is

Section titled “What a memory is”

A durable fact about the project: a curated fact, preference, decision or constraint any future session should recall. Every assistant in the project shares memories, whatever tool or conversation wrote them.

Memories are leads. A memory records what was true when it was written, so an assistant re-checks it against the code before acting, and removes it when it turns out wrong.

Where memories come from

Section titled “Where memories come from”
  • Assistants save them with remember when they learn something a later session could not recover.
  • Your other assistants contribute. Jensen reads the memories Claude Code, Codex and Hermes keep on your machine and merges them into project knowledge. Claude Code and its worktree memories are matched to the project. Codex and Hermes memories are global. It syncs about every 30 seconds.

The recall loop

Section titled “The recall loop”

Jensen nudges the learning loop at three points.

  1. Before the first edit, the assistant searches knowledge for the task and checks for a skill that fits.
  2. During work, a search hit that is a skill tells the assistant how to open it.
  3. At the end, when a session changed files or finished a plan and saved nothing, Jensen prompts a debrief. /jensen-debrief does the same by hand.

The knowledge settings

Section titled “The knowledge settings”

Settings, AI, Knowledge holds the store: statistics, the Embedding model, Refresh, Compact, Documentation sources with Index documentation, and the learning records waiting for you. While the model prepares, the panel reads Preparing semantic search.

What search covers

Section titled “What search covers”

One query reaches four sources: curated memories, recorded skills, ingested documentation and indexed source symbols. Assistants search it before reading files or guessing, which is where most of the time saving comes from.

You can search it too. Type a question in words into the command palette (Cmd K) and it searches by meaning.

It runs offline

Section titled “It runs offline”

Search is local. It combines a lexical index with a semantic one for ranking.

On first use Jensen tries to fetch a small embedding model to improve that ranking. With no network it falls back to lexical search. Nothing about search needs a remote service, and none of your code is sent anywhere to make it work.

Adding documentation

Section titled “Adding documentation”

Jensen indexes README files and local documentation folders beside the project and its services. Your git history stays untouched.

For a source it would not find, use Index documentation under Documentation sources. It takes a file or directory of markdown or plain text.

Review what was learned

Section titled “Review what was learned”

Not everything an assistant learns should guide the next one. Review learning records lists what is pending. Review with AI summarizes it, Approve keeps records that should guide future sessions, and Dismiss drops the rest.