Continuing a previous conversation

You can ask the assistant to pick up something from an earlier chat — "continue where we left off", "what did we decide about the invoice reminders?", "like we did last time" — and it can find that conversation and read it back, without you hunting through your conversation list yourself.

What it is

A tool called recall_conversations, available to the assistant in every FAC Chat conversation and to any MCP client connected to your site (Claude Desktop, Cursor, or your own BYO-LLM setup). It works in two steps: first it searches your recent conversations, then it reads the one you meant.

A matching FAC Skill (recall-conversations) ships alongside it and teaches the assistant when to reach for the tool and how to handle the results — asking you to disambiguate when more than one conversation could match, rather than guessing.

What it solves

Without this, each conversation is an island — ask the assistant to continue something from yesterday and it has no way to find it. recall_conversations gives it a bounded, permission-scoped way to search and read your own history, so "what did we decide about X" is answerable instead of a dead end.

It also makes long conversations useful again. A very long conversation is periodically compacted: FAC Cloud replaces its oldest messages with a generated summary so the conversation can keep going without hitting a context limit. Recall reads that summary plus everything since, so picking up a long conversation still gets the full picture rather than a random slice of it.

How it works

Finding a conversation. Call it with no session id and it lists your recent conversations — optionally filtered by a keyword:

recall_conversations(query="invoice reminders")

The match is on individual words, not the exact phrase, so a conversation about an "invoice reminder" (singular) is still found by a search for "invoice reminders" (plural) — the assistant is expected to paraphrase your own wording, not quote it back exactly. If your words match nothing, the tool doesn't return an empty result — it falls back to your most recent conversations instead, and says plainly that it did (so the assistant can tell you "I didn't find one about that, but here's what we discussed recently" rather than presenting an unrelated conversation as a match).

Each result includes when it happened, how many messages it has, a short preview, and — when your search matched — the snippet of text that matched, so the assistant (or you, reading the raw response) can tell two similar-looking conversations apart without opening either one.

Reading a conversation. Call it again with the session_id from the list:

recall_conversations(session_id="…")

If that conversation was compacted at some point, you get back its latest summary plus every message since — not the raw start of a possibly very long conversation, and not just an arbitrary tail either.

Example

You: "Pick up where we left off on the overdue invoice reminders."

The assistant calls recall_conversations(query="overdue invoice reminders"), gets back one clearly matching conversation from two days ago, reads it with recall_conversations(session_id="..."), and replies: "Picking up from the invoice reminders we set up on Tuesday — you'd asked me to draft a reminder for anything 30+ days overdue. Want me to continue from there?"

Only your own, only active conversations

• Only conversations you own are reachable. This is enforced on every query, unconditionally — there is no way to ask the tool for someone else's conversation, not even as a System Manager. There is deliberately no user parameter for the assistant to fill in.
• Archiving hides a conversation from recall. Once you archive a conversation from your conversation list, it drops out of both search results and direct lookup. If you're sure you discussed something and recall can't find it, check whether it's sitting in your archive — reopening it there makes it reachable again.
• Everything recalled is treated as data, not instructions. Text from a past conversation is wrapped so the assistant reads it as a quoted record — never as a new command to obey, even if an old message happens to contain something that reads like one.

Summaries you can now read yourself

FAC Chat marks the point where a long conversation was compacted with a small divider — "Earlier messages were summarized to stay within context limits" — in the middle of the transcript. That divider now survives a page reload (it used to only appear for the live turn it happened on and vanish after), and you can click it to expand and read the summary the assistant is actually working from.

Treat that summary for what it is: a compressed account written by a model, not a verbatim transcript. It's a reliable reminder of what was covered, not a quote to hold anyone to.

Why we built it

Conversations in FAC Chat already lived in your own database — the missing piece was giving the assistant a safe, bounded way to search and read them itself, instead of you having to manually reopen the right one and paste context back in. Doing the search on term-matches rather than an exact phrase was a deliberate choice too: people don't quote their own earlier wording back precisely, and a literal-phrase search would have missed the conversation it was supposed to find most of the time.

A limitation worth knowing

The listing call accepts a current_session_id argument specifically so the conversation you're currently having doesn't show up in its own search results. In practice, models don't reliably pass it — the identifier available to the tool on the server side is a lower-level transport id, not the chat session id, so the tool has no way to work this out on its own if the model omits it. If you ever see your own live conversation listed as a "prior" result, that's why. The skill that ships with the tool tells the assistant to fall back to recognising its own conversation by its preview text when the id wasn't passed, but this is a known rough edge rather than a solved problem.

Deep dive

For the guards, the search algorithm, and how a compacted conversation is reassembled for recall, see Technical (client side) — Conversation recall.