Recognise loose fact phrasings and reuse the wallet the asker just typed

README.md

@@ -219,19 +219,28 @@ chamar de Leo, nasci em 21/06/1981, meu telefone é …, moro em …, minha cart
 age is not kept: it follows from the birth date. "My name is" with one word is
 the name you're called by; with more, it is your full name. Common typos
 ("walet", "morro") and a missing "é"/"is" before an email, phone or wallet are
-understood.
+understood, and so is a request to remember wrapped around any of them —
+*"guarde na sua memória que eu quero os meus valores em reais brasileiro"*,
+*"lembre que …"*, *"remember that …"*.
 
-`what do you know about me?` (or `o que você sabe sobre mim?`) shows them,
-`what's my phone?` / `qual o meu telefone?` shows one, and `forget my email`
-deletes one. Replies come in the language you wrote in.
+`what do you know about me?` (or `o que você sabe sobre mim?`, also typed as
+`oq vc sabe sobre mim` or with a typo for "você") shows them, `what's my
+phone?` / `qual o meu telefone?` / `você sabe qual é o meu email?` shows one,
+and `forget my email` deletes one. `qual a moeda do meu país?` / `what's my
+currency?` shows the preferred currency, or says how to save one. Replies
+come in the language you wrote in.
 
 Saving another wallet adds it rather than replacing the first. `forget my
 wallet 0x…` (or `…45e0`) / `esqueça minha carteira 0x…` removes that one;
 with several saved, `forget my wallet` alone asks which. `forget my wallets` /
 `esqueça minhas carteiras` removes them all, Ethereum and Bitcoin. `what's my portfolio?`
 sums every saved wallet; a balance, DeFi, activity or alert question reads one, so with
 several saved it asks which (by their last four characters) unless you name it —
-`what's my wallet balance …45e0?`.
+`what's my wallet balance …45e0?`. With none saved and none typed, such a
+question uses the latest address you typed earlier in the same conversation (the
+same DM or channel — never one somebody else wrote or one found in the
+messages), and `a carteira que eu acabei de passar` / `essa carteira` / `the
+wallet I just gave you` asks your last wallet question again with it.
 
 With a preferred currency saved, every dollar figure the assistant reads for
 you is followed by the same figure in that currency, converted in code at the

openspec/changes/fix-facts-and-wallet-follow-up/proposal.md

@@ -0,0 +1,47 @@
+## Why
+
+One Portuguese DM in production hit four misses in a row:
+
+- "Guarde na sua memoria que eu quero os meus valores em reais brasileiro"
+  saved nothing: the "guarde (na sua memória) que" wrapper, "os meus valores"
+  and "reais brasileiro" were each unrecognised, so it went to the corpus.
+- "Oque vide sabe sobre mim ?" ("vide" for "você") went to the corpus.
+- "Você sabe qual é a moeda do meu país ?" went to the corpus.
+- After "verifique as pools de liquidez nesta wallet 0xB26B…45e0" was answered,
+  a follow-up about the value in the asker's currency was answered "Qual
+  carteira?", and "A carteira que eu acabei de passar" was searched for in the
+  corpus.
+
+## What Changes
+
+- **Remember wrappers.** "guarde / grave / anote / lembre / salve / registre
+  (na sua memória) que …" and "remember that …" before any fact statement.
+- **Currency phrasing.** "(eu) quero (ver) os (meus) valores em X", "I want to
+  see my prices in X"; the real in every number and gender ("reais
+  brasileiro", "real brasileiros", "brazilian reals").
+- **Show.** "o que / oque / oq <one word> sabe sobre mim" and "what do <one
+  word> know about me" when the word is "você" in any spelling or a short
+  lowercase typo of it; a capitalised name or another pronoun is not the
+  assistant. "Você sabe qual …" / "do you know what …" before a fact asked
+  for by name.
+- **The currency asked for.** "qual a moeda do meu país", "qual a minha moeda
+  de base", "what's the currency of my country", "which currency do I use"
+  show the preferred currency; with none saved the reply says so and how to
+  save one ("minha moeda é o real").
+- **The wallet just given.** A chain question (balance, positions, portfolio,
+  activity, alert) that needs an address and has none typed, carried or saved
+  reads the latest address the asker typed in this conversation -- their own
+  remembered questions, kept per person and location -- and logs its source
+  as `asker_typed`. A message that only points back at a wallet ("a carteira
+  que eu acabei de passar", "essa carteira", "the wallet I just gave you")
+  asks the latest chain question again with that address. The chain call
+  budget is charged per question as well as per address, so that follow-up
+  is not refused as the earlier run calling again.
+
+## Impact
+
+- `app/routing.py`, `app/routing_crypto.py`, `app/reasoning/service.py`,
+  `app/alert_intent.py`, `app/alert_requests.py`, `app/fact_replies.py`,
+  `domain/currency.py`, `adapters/chain/clearance.py`.
+- No schema change. Egress rooting is unchanged: the address is spelled into
+  the question, and it is the asker's own words.

openspec/changes/fix-facts-and-wallet-follow-up/specs/defi-positions/spec.md

@@ -0,0 +1,26 @@
+## ADDED Requirements
+
+### Requirement: A chain question without an address uses the one the asker just gave
+
+The system SHALL, when a chain question (balance, positions, portfolio,
+activity or alert) needs an address and none is typed, carried or saved, use
+the latest address the asker typed in their own remembered questions in the
+same conversation location, and SHALL never take one from another person's
+message, from retrieved content or from another location. A message that only
+points back at a wallet SHALL ask the asker's latest chain question again with
+that address.
+
+#### Scenario: A DM follow-up
+- WHEN, in a DM, "verifique as pools de liquidez nesta wallet 0x…" was answered
+  and the asker then writes "Porque você não respondeu o valor das pools na
+  minha moeda de base?" or "A carteira que eu acabei de passar"
+- THEN that address SHALL be read, and the reply SHALL NOT ask which wallet
+
+#### Scenario: A channel follow-up
+- WHEN the asker typed the address in the same channel conversation
+- THEN a follow-up there SHALL read it
+
+#### Scenario: Somebody else's address
+- WHEN an address appears only in another person's message, in retrieved
+  content, or in the asker's DM rather than this channel
+- THEN it SHALL NOT be read, and the reply SHALL ask for an address

openspec/changes/fix-facts-and-wallet-follow-up/specs/personal-facts/spec.md

@@ -0,0 +1,39 @@
+## ADDED Requirements
+
+### Requirement: Facts are recognised as people phrase them
+
+The system SHALL recognise a fact statement behind a request to remember it,
+the currency preference in its common phrasings, and a request to show the
+asker's facts with the assistant named in any common spelling or typo, without
+reading a question about money or about another person as either.
+
+#### Scenario: A currency kept in memory
+- WHEN someone writes "Guarde na sua memoria que eu quero os meus valores em
+  reais brasileiro"
+- THEN BRL SHALL be stored as their preferred currency
+
+#### Scenario: Money talk is not a preference
+- WHEN someone writes "quanto é 100 reais em dólar?" or "quero trocar 100
+  reais em dólar"
+- THEN no preferred currency SHALL be stored
+
+#### Scenario: "você" typed as people type it
+- WHEN someone writes "Oque vide sabe sobre mim ?"
+- THEN their facts SHALL be shown, as for "o que você sabe sobre mim?"
+
+#### Scenario: Somebody else is not the assistant
+- WHEN someone writes "o que o João sabe sobre mim?" or "o que ele sabe sobre
+  mim?"
+- THEN their facts SHALL NOT be shown for it
+
+### Requirement: The preferred currency can be asked for
+
+The system SHALL answer "qual a moeda do meu país", "qual a minha moeda",
+"what's my currency" and their variants with the preferred currency, and with
+none saved SHALL say so and how to save one, in the asker's language.
+
+#### Scenario: None saved
+- WHEN someone with no preferred currency writes "Você sabe qual é a moeda do
+  meu país ?"
+- THEN the reply SHALL say no preferred currency is saved and suggest "minha
+  moeda é o real", without searching the corpus

openspec/changes/fix-facts-and-wallet-follow-up/tasks.md

@@ -0,0 +1,19 @@
+## 1. Facts
+
+- [x] 1.1 Remember wrappers ("guarde na sua memória que", "lembre que", "anote que", "remember that")
+- [x] 1.2 "(eu) quero (ver) os (meus) valores em X"; the real in every number and gender
+- [x] 1.3 "o que <you or a typo> sabe sobre mim" / "what do <you> know about me"; names and pronouns excluded
+- [x] 1.4 "qual a moeda do meu país" and variants show the currency; none saved says how to save one
+
+## 2. The wallet just given
+
+- [x] 2.1 A chain route with no address uses the latest one the asker typed in this conversation
+- [x] 2.2 "a carteira que eu acabei de passar" / "essa carteira" / "the wallet I just gave you" re-asks the latest chain question
+- [x] 2.3 Alerts: the asker's earlier-typed address when nothing is typed, carried or saved
+- [x] 2.4 Chain call budget charged per question and address
+
+## 3. Tests and docs
+
+- [x] 3.1 Unit: the production sentences, false-positive guards, follow-up routing, budget
+- [x] 3.2 End to end: DM transcript, channel follow-up, another person's and the corpus's addresses never used
+- [x] 3.3 README facts and wallet sections

src/chatmemory/adapters/chain/clearance.py

@@ -144,8 +144,14 @@ async def _admit(
     key: str,
     cleared: Cleared,
 ) -> Cleared | ToolResult:
-    """The per-question budget, then the rate limit."""
-    if not budget.spend(key):
+    """The per-question budget, then the rate limit.
+
+    Charged per question as well as per address: a follow-up that reads the
+    wallet the asker just gave ("a carteira que eu acabei de passar") is a new
+    question, not the same run calling again. The rate limit still spaces
+    every call to the provider, whatever the question.
+    """
+    if not budget.spend(f"{cleared.question}\n{key}"):
         log.warning("chain.budget_exhausted", tool=tool)
         return refusal(server, tool, "calls_per_run_exhausted")
     if not await limiter.acquire():

src/chatmemory/app/alert_intent.py

@@ -65,7 +65,11 @@
 from dataclasses import dataclass
 from decimal import Decimal, InvalidOperation
 
-from chatmemory.app.routing import CONVERSATION_VERBS, recent_chain_address
+from chatmemory.app.routing import (
+    CONVERSATION_VERBS,
+    asker_typed_address,
+    recent_chain_address,
+)
 from chatmemory.domain.chain import find_addresses
 from chatmemory.ports.alerts import DEFAULT_EDGE_PERCENT, AlertKind, PriceDirection
 
@@ -267,6 +271,9 @@ class AlertIntent:
     asset: str | None = None
     direction: PriceDirection | None = None
     level: Decimal | None = None
+    #: With no address written or carried: the latest one the asker typed
+    #: earlier in this conversation, used only when no wallet is saved.
+    earlier_address: str | None = None
 
 
 def alert_intent(text: str, previous_questions: Sequence[str] = ()) -> AlertIntent | None:
@@ -289,6 +296,7 @@ def alert_intent(text: str, previous_questions: Sequence[str] = ()) -> AlertInte
         threshold=_threshold(text) if kind is AlertKind.AAVE_HEALTH else None,
         chain=_chain(text),
         address=address,
+        earlier_address=None if address else asker_typed_address(previous_questions),
         token_id=_token_id(text) if kind is AlertKind.LP_RANGE else None,
         edge_percent=_edge_percent(request) if kind is AlertKind.LP_RANGE else None,
     )

src/chatmemory/app/alert_requests.py

@@ -472,6 +472,9 @@ async def propose(
         if address is None and saved_wallets:
             wallets = suffix_list(tuple(saved_wallets))
             return AlertReply(text("which_wallet", language, wallets=wallets))
+        saved = address is not None and intent.address is None
+        # Typed by the asker earlier here, with nothing saved: theirs as well.
+        address = address or intent.earlier_address
         if address is None:
             return AlertReply(text("no_address", language))
         refused = _limit_refusal(intent, language)
@@ -482,7 +485,7 @@ async def propose(
         # An edge distance on a position already watched needs no new slot.
         if room <= 0 and intent.edge_percent is None:
             return AlertReply(text("at_cap", language, cap=self._cap))
-        source = AddressSource.TYPED if intent.address else AddressSource.SAVED
+        source = AddressSource.SAVED if saved else AddressSource.TYPED
         request = _Request(person, intent, address, source, language, direct)
         read = await self._targets.read(address, intent.kind, intent.chain)
         log.info(

src/chatmemory/app/fact_replies.py

@@ -230,6 +230,16 @@ def possessive(kind: FactKind, language: Language, *, plural: bool = False) -> s
         EN: "You haven't told me {p}.",
         PT: "Você ainda não me disse {p}.",
     },
+    "currency_missing": {
+        EN: (
+            "I don't have a preferred currency saved for you. Say "
+            "`my currency is the euro` and I'll show amounts in it too."
+        ),
+        PT: (
+            "Não tenho uma moeda preferida salva para você. Diga "
+            "`minha moeda é o real` e eu mostro os valores também nela."
+        ),
+    },
     "one_direct_only": {
         EN: "I only show {p} in a direct message. Ask me there.",
         PT: "Só mostro {p} em mensagem direta. Me pergunte lá.",
@@ -537,6 +547,9 @@ def _one_fact(
     # answer, not "you haven't told me".
     if kind is FactKind.FULL_NAME and FactKind.PREFERRED_NAME in by_kind:
         return shown_line(FactKind.PREFERRED_NAME, by_kind[FactKind.PREFERRED_NAME], language)
+    if kind is FactKind.PREFERRED_CURRENCY:
+        # Not "you haven't told me": how to tell it is the useful part.
+        return text("currency_missing", language)
     return text("one_missing", language, p=possessive(kind, language))
 
 

src/chatmemory/app/reasoning/service.py

@@ -76,6 +76,7 @@
     PositionKind,
     Route,
     RoutingDecision,
+    asker_typed_address,
     classify,
     explicit_web_search,
     market_follow_up,
@@ -401,6 +402,10 @@ async def _chain_route(
         """
         handler = CHAIN_HANDLERS[query.route]
         lookup = _lookup_addresses(question, query)
+        if query.asked is not None:
+            # "A carteira que eu acabei de passar": the earlier question, which
+            # is the asker's own words, is what is asked and rooted against.
+            question = replace(question, text=query.asked)
         addresses = lookup.addresses
         if lookup.choose_from:
             log.info(
@@ -435,6 +440,7 @@ async def _chain_route(
             route=handler.decision,
             label=str(query.route),
             carried=query.carried,
+            address_source=lookup.source,
             wallets=len(addresses),
         )
         return (
@@ -473,9 +479,33 @@ class _Lookup:
 
     addresses: tuple[str, ...] = ()
     choose_from: tuple[str, ...] = ()
+    #: Where the addresses came from, for the log: "typed" in this question,
+    #: "carried" from an earlier chain question, "asker_typed" earlier in this
+    #: conversation, "saved", or "" for none.
+    source: str = ""
 
 
 def _lookup_addresses(question: Question, query: CryptoQuery) -> _Lookup:
+    """Whose wallets to read, or which saved one to choose.
+
+    With nothing typed, carried or saved, the latest address the asker typed
+    in this conversation -- their own earlier words, recalled for them in
+    this location only, so another person's message or retrieved content is
+    never a source. "Which wallet?" was asked right after a question naming
+    one. The spelled-out address is then rooted in the asker's question like
+    any typed one.
+    """
+    found = _typed_or_saved(question, query)
+    if found.addresses or found.choose_from:
+        return found
+    previous = tuple(turn.question for turn in question.memory.turns)
+    earlier = asker_typed_address(previous)
+    if earlier is None:
+        return found
+    return _Lookup((earlier,), source="asker_typed")
+
+
+def _typed_or_saved(question: Question, query: CryptoQuery) -> _Lookup:
     """Whose wallets to read: what was typed or carried, and the saved ones.
 
     Every saved wallet joins a portfolio question about the asker's own money,
@@ -487,17 +517,19 @@ def _lookup_addresses(question: Question, query: CryptoQuery) -> _Lookup:
     otherwise asks which -- never silently the first.
     """
     typed = query.addresses
+    given = "carried" if query.carried else "typed"
     portfolio = query.route is CryptoRoute.PORTFOLIO
     if not query.mine or (typed and not portfolio):
-        return _Lookup(typed)
+        return _Lookup(typed, source=given if typed else "")
     saved = _saved_wallets(question)
     if not portfolio and len(saved) > 1:
         named = named_by_suffix(question.text, saved)
         if len(named) != 1:
             return _Lookup(choose_from=saved)
         saved = named
     lowered = {a.lower() for a in typed}
-    return _Lookup((*(w for w in saved if w.lower() not in lowered), *typed))
+    merged = (*(w for w in saved if w.lower() not in lowered), *typed)
+    return _Lookup(merged, source="saved" if saved else (given if typed else ""))
 
 
 def _spelled_out(question: Question, addresses: Sequence[str]) -> Question: