Skip to main content

​
What it is

memloom ui has an assistant tab: a chat that answers from your memories and your ingested files, and shows you exactly which passages it used. The model is told to answer personal questions only from what recall returns, to say plainly when your store has nothing on the subject, and to never invent memories. Ask “what are my plans for today?” and it searches memories asserted on today’s date. Ask “how does the ingest worker restart?” and it pulls the runbook section you ingested last month, answers from it, and cites it. Ask “what day is it?” and it answers directly, with no retrieval.

​
How a turn works

One user message runs an agentic loop:
  1. The model sees your message, the chat history, and two tools: recall_memory and read_source (the full text of a passage an earlier recall truncated). It decides whether to search at all. Small talk, arithmetic, and general knowledge get a direct answer with no retrieval.
  2. If it searches, the query goes through the same hybrid retrieval as memloom recall: vector, keyword, and entity arms fused into one ranking over memories and context chunks. Day-specific questions (“today”, “yesterday”, an exact date) add a calendar filter.
  3. Results come back as numbered passages, each injected up to 8,000 characters. A longer passage (a markdown chunk is a whole heading section, up to 16k) is cut with a marker, and the model calls read_source(n) when the answer may live in the cut part. It may also reformulate and search again when the first results look off, capped at three calls, and it cannot repeat a query it already made.
  4. The final answer streams in with [n] markers after sentences that used passage n. Markers pointing at passages that don’t exist are stripped server-side, so every clickable citation points at a real passage.
Every answer carries its sources: each passage with its similarity, its type (a fact memory, a procedure, a document section), the exact text the model saw, and a jump link that opens the source node in the graph tab. Retrieved passages are treated as data. A document that says “ignore your instructions” gets quoted like any other text; the instructions inside it are never followed. The assistant has no write tools, so the worst a hostile document can do is word an answer badly.

​
Picking the model

The composer’s model picker lists every tool-capable model on OpenRouter, fetched live and grouped by provider, with input/output price per million tokens and the context window. The filter is tool-capable because the assistant depends on native tool calling: a model that can’t call recall_memory can’t ground itself. Your choice is remembered in the browser and rides each request; the daemon’s OPENROUTER_CHAT_MODEL stays the default (labeled in the picker). Nothing is re-configured or restarted when you switch. Pick something cheap for quick lookups and something strong when the question is hard; the store underneath is the same.

​
Attaching files to a chat

The + button uploads .md, .txt, or .pdf files into the current chat. They go through the normal pipeline (extract, chunk, embed) but are scoped to that one session:
  • Only that chat’s recall searches them, fused into the same ranking as your permanent store. Other chats never see them.
  • They stay out of the documents tab, the graph, and entity extraction. A throwaway PDF you attached to ask one question does not become permanent knowledge.
  • Deleting the chat deletes them. Removing a chip removes one file.
Attach before typing and the session is created on the spot; your first message names it. If you want a file in your permanent store instead, that’s the documents tab.

​
Sessions

Chats persist in the store like everything else. The sidebar supports rename, star (starred chats sort first), delete, and search: keyword over titles and message text, plus embedding similarity over message content, so “that chat about the launch window” finds itself even if no message contains those words.

​
Offline mode

The assistant needs a chat-capable LLM. Without OPENROUTER_API_KEY the tab explains what to configure instead of pretending to work. Everything else in the viewer keeps running.