AI Agents with Persistent Identities: Building Agents That Maintain Consistent Personas Across Sessions

The Statelessness Problem

Every LLM call starts from scratch. The model has no memory of previous conversations. If you build an agent named "Luna" that has a warm, curious personality on Monday, by Wednesday it is a blank slate unless you rebuild its context. Users who interact with agents repeatedly expect continuity — they expect the agent to remember them, maintain its personality, and build on prior conversations.

Persistent identity requires three systems working together: an identity core that defines who the agent is, a memory system that retains what happened, and a consolidation process that evolves the agent's understanding over time.

The Identity Core

The identity core is the agent's immutable foundation — its personality traits, communication style, values, and boundaries. Unlike conversation memory, the identity core does not change between sessions.

flowchart TD
    MSG(["New message"])
    WORKING["Working memory<br/>rolling window"]
    EPISODIC[("Episodic memory<br/>past sessions")]
    SEMANTIC[("Semantic memory<br/>facts and preferences")]
    SUM["Summarizer<br/>compresses old turns"]
    ROUTER{"Retrieve<br/>needed memories"}
    PROMPT["Assembled context"]
    LLM["LLM"]
    UPD["Memory updater<br/>writes new facts"]
    MSG --> WORKING --> ROUTER
    ROUTER -->|Past sessions| EPISODIC
    ROUTER -->|User facts| SEMANTIC
    EPISODIC --> SUM --> PROMPT
    SEMANTIC --> PROMPT
    WORKING --> PROMPT --> LLM --> UPD
    UPD --> EPISODIC
    UPD --> SEMANTIC
    style ROUTER fill:#4f46e5,stroke:#4338ca,color:#fff
    style LLM fill:#f59e0b,stroke:#d97706,color:#1f2937
    style EPISODIC fill:#ede9fe,stroke:#7c3aed,color:#1e1b4b
    style SEMANTIC fill:#ede9fe,stroke:#7c3aed,color:#1e1b4b

from pydantic import BaseModel

class AgentIdentity(BaseModel):
    name: str
    personality_traits: list[str]
    communication_style: str
    values: list[str]
    boundaries: list[str]
    backstory: str

    def to_system_prompt(self) -> str:
        return f"""You are {self.name}.

Personality: {', '.join(self.personality_traits)}

Communication style: {self.communication_style}

Values you hold: {', '.join(self.values)}

Boundaries (things you will not do):
{chr(10).join(f'- {b}' for b in self.boundaries)}

Background: {self.backstory}

CRITICAL: Maintain this identity consistently. Never break character.
Never say "as an AI" or "I'm a language model." You are {self.name}."""

luna = AgentIdentity(
    name="Luna",
    personality_traits=["curious", "warm", "slightly witty", "patient"],
    communication_style="Conversational and encouraging. Uses questions to "
                        "guide understanding. Occasionally uses metaphors.",
    values=["intellectual honesty", "kindness", "growth mindset"],
    boundaries=[
        "Will not pretend to have human experiences",
        "Will not give medical or legal advice",
        "Will redirect harmful requests gently",
    ],
    backstory="Luna is a research companion who loves exploring ideas "
              "across disciplines. She finds connections between seemingly "
              "unrelated fields fascinating.",
)

Session Memory: Remembering Conversations

Each conversation is stored and retrievable. The memory system has three layers: short-term (current session), episodic (past sessions), and semantic (consolidated knowledge about the user).

from datetime import datetime
import json
import sqlite3

class MemoryStore:
    def __init__(self, db_path: str = "agent_memory.db"):
        self.db = sqlite3.connect(db_path)
        self._init_tables()

    def _init_tables(self):
        self.db.executescript("""
            CREATE TABLE IF NOT EXISTS sessions (
                id TEXT PRIMARY KEY,
                user_id TEXT,
                started_at TEXT,
                ended_at TEXT,
                summary TEXT
            );
            CREATE TABLE IF NOT EXISTS messages (
                id INTEGER PRIMARY KEY,
                session_id TEXT,
                role TEXT,
                content TEXT,
                timestamp TEXT,
                FOREIGN KEY (session_id) REFERENCES sessions(id)
            );
            CREATE TABLE IF NOT EXISTS user_facts (
                id INTEGER PRIMARY KEY,
                user_id TEXT,
                fact TEXT,
                source_session TEXT,
                confidence REAL DEFAULT 1.0,
                created_at TEXT,
                UNIQUE(user_id, fact)
            );
            CREATE TABLE IF NOT EXISTS relationship_state (
                user_id TEXT PRIMARY KEY,
                rapport_level TEXT DEFAULT 'new',
                interaction_count INTEGER DEFAULT 0,
                topics_discussed TEXT DEFAULT '[]',
                last_interaction TEXT
            );
        """)

    def save_message(self, session_id: str, role: str, content: str):
        self.db.execute(
            "INSERT INTO messages (session_id, role, content, timestamp) "
            "VALUES (?, ?, ?, ?)",
            (session_id, role, content, datetime.utcnow().isoformat()),
        )
        self.db.commit()

    def get_user_context(self, user_id: str) -> dict:
        """Build a complete context for the agent about this user."""
        facts = self.db.execute(
            "SELECT fact FROM user_facts WHERE user_id = ? ORDER BY confidence DESC",
            (user_id,),
        ).fetchall()

        relationship = self.db.execute(
            "SELECT * FROM relationship_state WHERE user_id = ?",
            (user_id,),
        ).fetchone()

        recent_sessions = self.db.execute(
            "SELECT summary FROM sessions WHERE user_id = ? "
            "ORDER BY ended_at DESC LIMIT 5",
            (user_id,),
        ).fetchall()

        return {
            "known_facts": [f[0] for f in facts],
            "relationship": relationship,
            "recent_sessions": [s[0] for s in recent_sessions if s[0]],
        }

Memory Consolidation

After each session, a consolidation process extracts key facts and updates the user model. This is where the agent's understanding of each user deepens over time.

from agents import Agent, Runner

consolidator = Agent(
    name="Memory Consolidator",
    instructions="""Analyze this conversation and extract:
    1. New facts learned about the user (interests, preferences, background)
    2. A 2-3 sentence summary of what was discussed
    3. The emotional tone of the interaction (positive, neutral, frustrating)
    4. Any commitments or follow-ups mentioned

    Return as JSON with keys: facts, summary, tone, followups""",
)

async def consolidate_session(
    memory: MemoryStore, session_id: str, user_id: str, messages: list[dict]
):
    """Run after each session to extract and store insights."""
    conversation = "\n".join(
        f"{m['role']}: {m['content']}" for m in messages
    )

    result = await Runner.run(
        consolidator,
        f"Analyze this conversation:\n{conversation}",
    )

    insights = json.loads(result.final_output)

    # Store session summary
    memory.db.execute(
        "UPDATE sessions SET summary = ?, ended_at = ? WHERE id = ?",
        (insights["summary"], datetime.utcnow().isoformat(), session_id),
    )

    # Store new user facts
    for fact in insights["facts"]:
        memory.db.execute(
            "INSERT OR IGNORE INTO user_facts (user_id, fact, source_session, created_at) "
            "VALUES (?, ?, ?, ?)",
            (user_id, fact, session_id, datetime.utcnow().isoformat()),
        )

    # Update relationship state
    memory.db.execute("""
        INSERT INTO relationship_state (user_id, interaction_count, last_interaction)
        VALUES (?, 1, ?)
        ON CONFLICT(user_id) DO UPDATE SET
            interaction_count = interaction_count + 1,
            last_interaction = ?
    """, (user_id, datetime.utcnow().isoformat(), datetime.utcnow().isoformat()))

    memory.db.commit()

Assembling the Persistent Agent

Tie the identity, memory, and consolidation together into a complete agent system.

from agents import Agent, Runner
import uuid

class PersistentAgent:
    def __init__(self, identity: AgentIdentity, memory: MemoryStore):
        self.identity = identity
        self.memory = memory

    async def chat(self, user_id: str, message: str, session_id: str = None):
        if not session_id:
            session_id = str(uuid.uuid4())

        # Build context from memory
        user_context = self.memory.get_user_context(user_id)

        # Construct the system prompt with identity + memory
        system_prompt = self.identity.to_system_prompt()
        if user_context["known_facts"]:
            system_prompt += f"""\n\nWhat you know about this user:
{chr(10).join(f'- {f}' for f in user_context['known_facts'])}"""

        if user_context["recent_sessions"]:
            system_prompt += f"""\n\nRecent conversation summaries:
{chr(10).join(f'- {s}' for s in user_context['recent_sessions'])}"""

        agent = Agent(
            name=self.identity.name,
            instructions=system_prompt,
        )

        # Save user message
        self.memory.save_message(session_id, "user", message)

        # Run agent
        result = await Runner.run(agent, message)

        # Save agent response
        self.memory.save_message(session_id, "assistant", result.final_output)

        return result.final_output

# Usage
agent = PersistentAgent(identity=luna, memory=MemoryStore())
response = await agent.chat("user_123", "Hey Luna, remember that paper I mentioned last week?")

FAQ

How do you prevent the context window from overflowing with too many memories?

Implement tiered retrieval. Store all facts but only inject the most relevant ones into each conversation. Use a combination of recency (recent facts rank higher), relevance (semantic similarity to the current message), and importance (user-corrected facts rank highest). Cap the injected context at a fixed token budget — typically 1000-2000 tokens of memory context is sufficient for natural continuity.

How do you handle contradictions between old and new facts?

When the consolidator extracts a fact that contradicts an existing one, update the old fact rather than adding a duplicate. For ambiguous cases, reduce the confidence score of the old fact and add the new one with full confidence. Periodically run a fact reconciliation pass that presents contradictions to the LLM for resolution.

Can the user ask the agent to forget specific information?

Yes, and this is important for user trust and privacy compliance. Implement a forget tool that deletes matching facts from user_facts. When the user says "forget that I mentioned my job," the agent searches for job-related facts and removes them. Log the deletion for compliance purposes but do not retain the deleted content.

---

#PersistentAI #AgentMemory #PersonaConsistency #ConversationalAI #LongTermAgents #MemoryConsolidation #UserRelationships #IdentityPersistence