Implement session-based semantic memory and retrieval

> Currently, TopSecret-Agent stores individual conversation turns in SQLite, but it does not have a mechanism for retrieving relevant information from previous sessions.
>
> Implement a session-based semantic memory system where each completed session is summarized and indexed using an embedding. When a new user query arrives, the system should retrieve relevant sessions using semantic similarity and then fetch the complete conversation for those sessions from SQLite.

### Proposed architecture

```text
User Query
│
▼
Generate Query Embedding
│
▼
Vector Search
│
▼
Relevant Session IDs
│
▼
SQLite
│
▼
Full Session Conversations
│
▼
Agent Prompt
```

### Session indexing

When a session ends:

```text
Conversation
│
▼
Generate summary
│
▼
Generate embedding
│
▼
Store embedding + session_id
```

SQLite remains the source of truth for the actual conversation.

### Requirements

* [ ] Introduce an explicit session concept.
* [ ] Associate every conversation turn with a session ID.
* [ ] Generate a summary for a completed session.
* [ ] Generate an embedding from the session summary.
* [ ] Store the embedding together with the session ID.
* [ ] Generate an embedding for new user queries.
* [ ] Retrieve the most relevant session IDs.
* [ ] Fetch the corresponding conversations from SQLite.
* [ ] Provide retrieved sessions to the agent as additional context.
* [ ] Keep the existing user-profile mechanism separate from episodic memory.

### Design principle

The vector store should act as a **retrieval index**, not the source of truth.

```text
Vector Store
↓
"Session 17 looks relevant"
↓
SQLite
↓
"Here is everything that happened in Session 17"
```

### Out of scope

For the first implementation:

* No message-level embeddings
* No automatic memory deletion
* No complex memory ranking
* No hybrid BM25 + vector retrieval
* No sophisticated long-term memory management

Those can be separate issues later.