Memory page, and showing what memory did in chat - #261
Open
farhat-is-coding wants to merge 12 commits into
Open
Conversation
Editing is no longer a disabled pencil with a promise attached — the layered backend is here, so cards and USER.md lines are rewritten in place through a shared InlineEditor (Enter saves, Escape cancels). The field stays open while the save is in flight and when the server refuses, so a rejected edit — a rule whose new trigger collides, a line that would cross the token ceiling — keeps the text there to fix rather than throwing the work away. Sample mode is gone entirely: memorySamples.ts, previewMode, the seed endpoint and thunk, the local-only delete path, the sample banner and empty -state button, and the semanticDisabled prop that existed only to hobble search while sample data was showing. Everything on the page is now the user's real store. Also drops the permanently-disabled whole-file pencil (per-line editing replaces it) and corrects the explainer's claim that the writer and gate are still coming — only consolidation is.
The switch had no explanation attached, but it is the one control in this panel that both reads and writes: on means DARE brings what it knows into the conversation AND reads each finished reply for anything worth keeping. That is a consent decision, so the panel now states it in those terms and flips the wording with the switch. Adds the route the panel was missing — 'See what it remembers' opens the Memory page, so the toggle and the store it fills are one click apart. Also swaps the Brain icon for the Fingerprint the Memory page already uses, and labels the switch for screen readers.
Two things the store knew and the interface never said. A turn's memory activity now ends with what the writer decided, on the same timeline as the tools and sources — including the decisions it refused, since those are the ones that show the gate working. Retired memories get their own disclosure below the feed, each paired with what replaced it: Pittsburgh struck through, Boston beside it. They stay out of the main feed on purpose — mixed in with current memories they read as contradictions, with nothing on the card saying which one is live. Also swaps the recalled-memories icon from Brain to Fingerprint, matching the Memory page.
The toggle now switches transcript search too, so the panel says so.
docs/RULES.md: one component per file, helpers included. MemoryArchive held RetiredRow and MemoryWriteStep held ChangeRow.
It sat on the same timeline rail as the tool calls, with the same icon treatment, so it read as a step the turn had taken. It isn't one — the writer reads the finished turn in a background job, seconds later, in another process. It now sits below the rail under "After the reply". And it led with process instead of outcome. One "call me Abbas" produced three ledger entries as the gate worked through a patch and two supersedes, each describing a state that no longer exists by the time you read it. The headline is now the last applied decision per memory — what the store actually holds — with every decision, refusals included, one disclosure away.
The sweep finds duplicates, facts repeated enough to belong in the profile, slots named for what they used to hold, and a profile with more pinned to it than fits. Each one is a card that leads with the memory rather than the verb, because the question being asked is "is this right about you?" — answered by reading the sentence, not by parsing what "merge" means. Opt-in rather than always-on. A panel permanently reading "6 things need your attention" is a chore; asked for, it either has something to say or says plainly that it does not. Both buttons are safe: approving retires rather than deletes, and passing changes nothing. Approving refetches the store so the archive count and the layer cards move with it. Dismissal lives in the panel rather than in each card. Held per-card, the last "Leave it" emptied the panel with no closing state — it looked broken instead of finished.
…page The Sessions card used to open the explainer — a dead end dressed as a layer. It now selects like the other three, and switches the search surface below into transcript mode: the same search the model reaches through its search_sessions tool, so what the page can find and what the model can find are one and the same. Words, a date range, or both (dates alone answer "what did we talk about last week?"); each hit shows the matched exchange with the turn either side and opens the conversation it came from. State lives in the memory slice — one search surface, two modes, entering either clears the other's results. Also rewrites the How-it-works sheet against what shipped: the recall step now says what actually happens (profile whole, best 3 facts by meaning, up to 5 rules, a relevance bar), the separate cold-layer funnel section is gone, consolidation is described as the live Tidy up sweep rather than a thing rolling out next, editing is no longer 'on the way', and the FTS5 label — SQLite's module, which this never was — reads FTS.
Export downloads the dare-memory-v2 bundle — the layered contract itself, not a flat list. Import takes a bundle by file or paste, shows what is inside before anything happens (active / retired / held / rules, document or not), and surfaces the backend's refusals verbatim — the empty-store rule above all. Successor to the MemU-era export dialog, whose portability idea survives here with a format that can actually reinstate a store.
Second mode in the Import dialog. Step one hands the person a prompt for their other assistant — plain bullet points in their own voice, exact wording preserved for instructions, nothing session-specific. Step two takes whatever comes back and queues it through DARE's own memory pipeline; the toast says work has started rather than pretending it finished, because the writer works through the queue.
3 tasks
Memory page: export, import, and import from another AI
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
The interface for the layered memory backend (cmudco/dare-backend#338).
Memory page — four layer cards, semantic search with relevance meters, inline editing, and "What DARE used to think": retired memories paired with whatever replaced them, so a correction reads as a correction rather than as data loss.
In chat — the per-message activity panel now shows the write side, not just the read side. What was remembered, what was retired, and what the gate refused, each with the writer's own reason. A decision the gate overruled says what was asked for alongside what was done.
The memory toggle now means what its copy says: it gates retrieval, the writer, and transcript search. The duplicate
search_sessionscheckbox is gone from the tool drawer.Notable
Test plan
npm run app:buildclean, 0 lint errors