main character

One chapter stays open for as long as it feels right, which is about a week in practice. The write screen is that whole conversation: the chapter it continues, your entries, the companion's replies, and everything you say back. save entry stores a journal entry (backed up to a markdown file the moment you save), and send talks with the companion about it. Both join the same chapter and both become journal memory at close. Your draft autosaves as you type and survives a refresh or closed tab. Enter just makes a new line, and only the buttons send.

When the chapter feels finished, open the ⋯ menu next to the composer and hit close chapter (also available on the history tab). That's the summarize point. Your side of the chapter becomes a journal entry, the same shape as your imported conversations, and tagging, entity extraction, summaries, and dream scanning all run in the background. A new, empty chapter opens. Until then, new entries live in the conversation (and their markdown backups), not yet in long-term memory. Once your own writing in a chapter passes about 30,000 characters (Settings changes the figure), the write screen asks whether you're ready to close. Say not yet, and it asks again after another 10,000.

Dreams need no flag. Write them the way you always have ("last night I had a dream…") and the dream extractor finds them at the summarize point, pulling each one into the dream realm while the entry itself stays in waking memory. That's exactly how your imported history behaves.

The chat screen is a tool, not the journal. Ask it to pull things out of your history ("when did I last see Sam?", "what did I say about the move?"). It runs the same retrieval as the companion, with its own finder prompt, but it's a separate conversation, so lookups never join the open chapter and never become journal memory. clear wipes it whenever you like. Its honest limit is that it sees retrieved excerpts, not the whole journal, so answers are representative rather than exhaustive.

smart replies, the toggle by send (off by default), lets it search again on its own before answering: other words, exact words, a stretch of dates, a whole entry. It is much better at "when did I first…", "how often…" and how something changed. Each question can take several calls, so it costs more and takes a few seconds longer.

The search screen sweeps your entire journal locally, so it's free and instant with no API calls. meaning mode ranks passages by how semantically close they are to whatever you describe, in your own words ("times I felt proud of my work"), while exact text mode finds literal matches, newest first, with a ×N count when an entry contains several. Click any result to read the full entry in place. This is the exhaustive counterpart to the chat screen. Chat retrieves a handful of excerpts and reasons about them, while search shows you everything. Dreams never appear here because they live in their own realm (dreams tab).

There are two honest limits. Highlighting is literal (a typo'd word won't light up even though meaning-search still finds the right passages), and dates are the conversation's date, so a passage written mid-week inside a week-long chapter shows that chapter's date until the per-day split lands.

Lands on the open chapter: the conversation it continues (with the companion's replies interleaved, backfilled from your export), your entries, and the live thread. Every closed chapter sits in the list, and you can pick one to read it in full. Closed chapters keep both sides for reading, but only your side is journal memory.

Every entry gets tagged with categories (work, relationships, health…) automatically when a chapter closes, and tag new entries re-runs it manually for anything untagged (already-tagged entries are cached and free). Pick a category to browse its entries and read its domain document, a living ~500-word summary of that area of life. × a wrong tag or + a missing one, and your fixes become permanent overrides. scan for new categories clusters your places/projects and proposes organic categories, but nothing is created without your confirm. You can also hand-seed a category with keywords. Entry categories are deliberately flat to stay broad and scannable, so there's no sub-category nesting here.

detect patterns re-reads the weekly arcs + domain documents and names recurring shapes (emotional cycles, behavioral pipelines, relationship dynamics), each with dated instances and a confidence level. Patterns are injected into the companion's context only when the conversation genuinely rhymes with one. dismiss hides a pattern until new evidence shows up.

Dreams live in their own realm, a separate collection that waking questions can never accidentally surface. The extractor finds dreams inside your entries (at the summarize point, or via extract dreams from history), keeping the narrative, cast, and tones, plus your own interpretation if you gave one. Recent dream tone shows up as the dream weather line. Asking the companion about dreams crosses the boundary explicitly.

Everything is undoable, since ↺/↻ on the entities screen cover every curation and observation change, 50 steps back. All your decisions live in entity_graph/curation.json and re-apply on every rebuild, so nothing you do here is ever lost by the system regenerating docs. The one exception is a --force re-extraction (a from-scratch API re-read of all entries, only ever run deliberately), which rebuilds the raw memory files and would discard observation-level edits. Merges, renames and the rest survive even that. Removed aliases keep their historical "as ‘name’" data, because removal only stops name-matching.

Your journal is just the local folders chroma_data, journal_entries, entity_graph and summaries, so back them up.