CLI Chatbot with Memory

 terminal (prompt_toolkit / input)        │        ▼ ┌──────────────┐   build context    ┌───────────────────┐ │  REPL loop   │ ─────────────────► │ ContextBuilder    │ │  + commands  │                    │  system prompt    │ └──────┬───────┘                    │  + memory facts   │        │                            │  + rolling summary│        │                            │  + recent turns   │        │                            └─────────┬─────────┘        │ stream                               │ messages        ▼                                      ▼ ┌──────────────┐                    ┌───────────────────┐ │ LLM adapter  │ ◄───────────────── │  token counter    │ │ (stream API) │                    └───────────────────┘ └──────┬───────┘        │ reply + usage        ▼ ┌──────────────────────────────────────────────┐ │ SQLite: sessions, messages, summaries, facts, usage │ └──────────────────────────────────────────────┘        ▲        └── background: summarize overflow, extract facts (cheap model)

import sqlite3, jsonfrom dataclasses import dataclass@dataclassclass Msg:    role: str    content: strclass ContextBuilder:    def __init__(self, db: sqlite3.Connection, llm, count_tokens, budget=6000, summary_cap=600):        self.db, self.llm, self.count = db, llm, count_tokens        self.budget, self.summary_cap = budget, summary_cap    def build(self, session_id: str, system: str) -> list[dict]:        turns = [Msg(r, c) for r, c in self.db.execute(            "SELECT role, content FROM messages WHERE session_id=? AND folded=0 ORDER BY id",            (session_id,))]        summary = self._summary(session_id)        facts = self._facts(limit=10)        header = system + ("\n\nKnown about the user:\n- " + "\n- ".join(facts) if facts else "")        used = self.count(header) + self.count(summary)        kept: list[Msg] = []        for m in reversed(turns):            t = self.count(m.content)            if used + t > self.budget:                break            kept.insert(0, m); used += t        overflow = turns[: len(turns) - len(kept)]        if overflow:            summary = self._fold(session_id, summary, overflow)        msgs = [{"role": "system", "content": header}]        if summary:            msgs.append({"role": "system", "content": f"Earlier conversation summary:\n{summary}"})        return msgs + [{"role": m.role, "content": m.content} for m in kept]    def _fold(self, session_id, summary, overflow):        text = "\n".join(f"{m.role}: {m.content}" for m in overflow)        new = self.llm.complete(model="small", max_tokens=self.summary_cap, messages=[            {"role": "system", "content": "Update the running summary. Keep names, decisions, "             "open questions, and user preferences. Be terse."},            {"role": "user", "content": f"Current summary:\n{summary}\n\nNew turns:\n{text}"}]).text        self.db.execute("UPDATE messages SET folded=1 WHERE session_id=? AND folded=0 AND id IN "                        "(SELECT id FROM messages WHERE session_id=? AND folded=0 ORDER BY id LIMIT ?)",                        (session_id, session_id, len(overflow)))        self.db.execute("INSERT OR REPLACE INTO summaries(session_id, text) VALUES (?, ?)",                        (session_id, new))        self.db.commit()        return new    def _summary(self, sid):        row = self.db.execute("SELECT text FROM summaries WHERE session_id=?", (sid,)).fetchone()        return row[0] if row else ""    def _facts(self, limit):        return [r[0] for r in self.db.execute(            "SELECT text FROM facts ORDER BY updated_at DESC LIMIT ?", (limit,))]