Retrieval at Query Time

The Retrieval Pipeline

public class SemanticMemoryService : ISemanticMemoryService
{
    public async Task<SemanticRetrievalResult> RetrieveAsync(
        AgentComposite agent,
        string query,
        CancellationToken ct = default)
    {
        if (!agent.MemoryConfig.SemanticEnabled)
            return SemanticRetrievalResult.Empty;

        // 1. Embed the query
        var queryEmbedding = await _embeddingProvider.EmbedAsync(query, ct);

        // 2. Search the agent's vector collection
        var collection = GetCollectionName(agent);
        var results = await _store.SearchAsync(
            collection,
            queryEmbedding,
            topK: agent.MemoryConfig.SemanticTopK,
            minScore: agent.MemoryConfig.SemanticMinScore,
            filter: new MemoryFilter { TenantId = agent.TenantId },
            ct: ct);

        // 3. Format for context injection
        return new SemanticRetrievalResult
        {
            Chunks = results,
            FormattedContext = FormatForContext(results)
        };
    }

    private string FormatForContext(IReadOnlyList<MemoryRecord> records)
    {
        var sb = new StringBuilder("[Retrieved Knowledge]\n");
        foreach (var r in records)
        {
            sb.AppendLine($"Source: {r.Metadata.Source}");
            sb.AppendLine(r.Content);
            sb.AppendLine("---");
        }
        return sb.ToString();
    }
}

Retrieval Configuration

Context Injection Position

Retrieved knowledge is injected between the system prompt and the conversation history in the LLM context:

// LLM context assembly order:
[1] System Prompt        ← agent.SystemPrompt
[2] Retrieved Knowledge  ← semantic retrieval results (this page)
[3] Episode Snippets     ← episodic memory recall
[4] Message History      ← current session messages (pruned to budget)
[5] Current User Message ← the user's latest input

// The LLM reads [2] Retrieved Knowledge as authoritative information
// to ground its response — reducing hallucination