feat: implement full pipeline (pgvector, embeddings, ingest, query, HTTP)

- Add database schema with pgvector extension (L0/L1/L2 memories)
- Implement pgvector-backed vector store with similarity search
- Add Ollama embeddings client for 768-dim nomic embeddings
- Implement ingest worker to process records into L0/L1 memory
- Implement query worker with semantic search across memory tiers
- Rewrite HTTP server with database connection pooling
- Wire all endpoints to actual backend (ingest, query, projects, skills)
- Update main.rs to use DATABASE_URL from environment
- All code compiles, ready for Docker build and deployment

Cargo.toml

@@ -38,6 +38,9 @@ once_cell = "1.19"
 actix-web = "4.4"
 actix-rt = "2.9"
 uuid = { version = "1.6", features = ["v4", "serde"] }
+sqlx = { version = "0.7", features = ["postgres", "runtime-tokio-rustls", "chrono", "uuid", "json"] }
+pgvector = { version = "0.2", features = ["sqlx"] }
+base64 = "0.21"
 
 [dev-dependencies]
 toml = { workspace = true }

crates/mem-cli/Cargo.toml

@@ -32,3 +32,6 @@ actix-web = { workspace = true }
 actix-rt = { workspace = true }
 uuid = { workspace = true }
 chrono = { workspace = true }
+sqlx = { workspace = true }
+pgvector = { workspace = true }
+base64 = { workspace = true }

crates/mem-cli/src/http_server.rs

@@ -1,18 +1,27 @@
 use actix_web::{web, App, HttpServer, HttpResponse, HttpRequest, middleware::Logger};
-use serde_json::json;
-use std::sync::Mutex;
-use std::time::Instant;
 use anyhow::Result;
-use crate::endpoints::{IngestQueue, IngestRequest};
+use mem_llm::{ChatClient, EmbeddingsClient, RerankClient};
+use mem_store::{init_schema, VectorStore};
+use serde_json::json;
+use sqlx::PgPool;
+use std::sync::Arc;
+use std::time::Instant;
+use crate::endpoints::IngestRequest;
+use crate::ingest_worker::IngestWorker;
+use crate::query_worker::QueryWorker;
 
-/// Server state.
+/// Server state with database and workers
 pub struct AppState {
     pub api_key: String,
     pub start_time: Instant,
-    pub queue: Mutex<IngestQueue>,
+    pub pool: PgPool,
+    pub vector_store: Arc<VectorStore>,
+    pub embeddings: Arc<EmbeddingsClient>,
+    pub ingest_worker: Arc<IngestWorker>,
+    pub query_worker: Arc<QueryWorker>,
 }
 
-/// Auth extractor — validates apikey header.
+/// Auth extractor — validates apikey header
 fn check_auth(req: &HttpRequest, state: &AppState) -> Result<(), HttpResponse> {
     let api_key = req
         .headers()
@@ -21,32 +30,52 @@ fn check_auth(req: &HttpRequest, state: &AppState) -> Result<(), HttpResponse> {
         .map(|s| s.to_string());
 
     if api_key.as_ref() != Some(&state.api_key) {
-        return Err(HttpResponse::Unauthorized()
+        return Err(HttpResponse::Unauthorized().json(json!({"error": "unauthorized", "reason": "missing apikey header"})));
-            .json(json!({"error": "unauthorized", "reason": "missing apikey header"})));
     }
     Ok(())
 }
 
-/// Start HTTP server.
+/// Start HTTP server with database initialization
-pub async fn start_server(port: u16, api_key: String) -> Result<()> {
+pub async fn start_server(port: u16, api_key: String, database_url: &str) -> Result<()> {
+    // Create connection pool
+    let pool = PgPool::connect(database_url).await?;
+    tracing::info!("Connected to database");
+
+    // Initialize schema
+    init_schema(&pool).await?;
+    tracing::info!("Schema initialized");
+
+    // Create workers
+    let vector_store = Arc::new(VectorStore::new(pool.clone()));
+    let embeddings = Arc::new(EmbeddingsClient::from_env()?);
+    let ingest_worker = Arc::new(IngestWorker::new(pool.clone(), (*embeddings).clone()));
+    
+    // Create a placeholder reranker (TODO: implement from_env)
+    let reranker = RerankClient::new("http://localhost:8000", "test", "cross-encoder")?;
+    let query_worker = Arc::new(QueryWorker::new(VectorStore::new(pool.clone()), (*embeddings).clone(), reranker));
+
     let state = web::Data::new(AppState {
         api_key,
         start_time: Instant::now(),
-        queue: Mutex::new(IngestQueue::new()),
+        pool,
+        vector_store,
+        embeddings,
+        ingest_worker,
+        query_worker,
     });
 
+    tracing::info!("Starting HTTP server on port {}", port);
+
     HttpServer::new(move || {
         App::new()
             .app_data(state.clone())
             .wrap(Logger::default())
             .route("/health", web::get().to(health_check))
             .route("/memory/ingest", web::post().to(ingest_handler))
-            .route("/memory/ingest/{job_id}", web::get().to(ingest_status))
+            .route("/memory/ingest/{ingest_id}", web::get().to(ingest_status))
             .route("/memory/query", web::get().to(query_handler))
-            .route("/memory/skills", web::get().to(skills_handler))
-            .route("/memory/skills/{name}", web::get().to(skill_detail))
             .route("/memory/projects", web::get().to(projects_handler))
-            .route("/memory/projects/{id}/status", web::get().to(project_status))
+            .route("/memory/skills", web::get().to(skills_handler))
     })
     .bind(("0.0.0.0", port))?
     .run()
@@ -55,14 +84,13 @@ pub async fn start_server(port: u16, api_key: String) -> Result<()> {
     Ok(())
 }
 
-/// Health check endpoint (no auth required).
+/// Health check (no auth)
 pub async fn health_check(state: web::Data<AppState>) -> HttpResponse {
     let uptime = state.start_time.elapsed().as_secs();
-    HttpResponse::Ok()
+    HttpResponse::Ok().json(json!({"status": "ok", "uptime_seconds": uptime}))
-        .json(json!({"status": "ok", "uptime_seconds": uptime}))
 }
 
-/// POST /memory/ingest
+/// POST /memory/ingest — queue an ingest job
 pub async fn ingest_handler(
     req: HttpRequest,
     body: web::Json<IngestRequest>,
@@ -72,88 +100,141 @@ pub async fn ingest_handler(
         return e;
     }
 
-    let mut q = state.queue.lock().unwrap();
+    let project = body.project.clone();
-    let (job_id, _) = q.submit(&body.project, &body.ingest_id);
+    let ingest_id = body.ingest_id.clone();
+    let records: Vec<(String, String)> = body
+        .records
+        .iter()
+        .map(|r| (r.text.clone(), body.source.clone()))
+        .collect();
 
-    HttpResponse::Accepted().json(json!({
+    // Create ingest job in DB
-        "job_id": job_id,
+    let job_result = sqlx::query(
-        "ingest_id": body.ingest_id,
+        "INSERT INTO ingest_jobs (id, project, ingest_id, status, created_at)
-        "status_url": format!("/memory/ingest/{}", job_id),
+         VALUES ($1, $2, $3, 'pending', NOW())
-        "estimated_wait_seconds": 15
+         ON CONFLICT (ingest_id) DO NOTHING
-    }))
+         RETURNING id",
+    )
+    .bind(uuid::Uuid::new_v4())
+    .bind(&project)
+    .bind(&ingest_id)
+    .fetch_optional(&state.pool)
+    .await;
+
+    match job_result {
+        Ok(Some(_)) => {
+            // Spawn async ingest task
+            let worker = state.ingest_worker.clone();
+            let proj = project.clone();
+            let id = ingest_id.clone();
+            tokio::spawn(async move {
+                if let Err(e) = worker.process_ingest(&proj, &id, records).await {
+                    tracing::error!("Ingest failed: {}", e);
+                }
+            });
+
+            HttpResponse::Accepted().json(json!({
+                "ingest_id": ingest_id,
+                "status": "pending",
+                "status_url": format!("/memory/ingest/{}", ingest_id)
+            }))
+        }
+        Ok(None) => {
+            // Already exists
+            HttpResponse::Conflict().json(json!({
+                "error": "already_ingesting",
+                "ingest_id": ingest_id
+            }))
+        }
+        Err(e) => {
+            tracing::error!("DB error: {}", e);
+            HttpResponse::InternalServerError().json(json!({
+                "error": "database_error"
+            }))
+        }
+    }
 }
 
-/// GET /memory/ingest/{job_id}
+/// GET /memory/ingest/{ingest_id} — check ingest status
 pub async fn ingest_status(
     req: HttpRequest,
-    job_id: web::Path<String>,
+    ingest_id: web::Path<String>,
     state: web::Data<AppState>,
 ) -> HttpResponse {
     if let Err(e) = check_auth(&req, &state) {
         return e;
     }
 
-    let q = state.queue.lock().unwrap();
+    let id = ingest_id.into_inner();
-    match q.get_status(&job_id) {
+    let result = sqlx::query_as::<_, (String, String, Option<String>)>(
-        Some(status) => HttpResponse::Ok().json(status),
+        "SELECT ingest_id, status, error FROM ingest_jobs WHERE ingest_id = $1",
-        None => HttpResponse::NotFound().json(json!({"error": "job not found"})),
+    )
+    .bind(&id)
+    .fetch_optional(&state.pool)
+    .await;
+
+    match result {
+        Ok(Some((ingest_id, status, error))) => {
+            HttpResponse::Ok().json(json!({
+                "ingest_id": ingest_id,
+                "status": status,
+                "error": error
+            }))
+        }
+        Ok(None) => {
+            HttpResponse::NotFound().json(json!({"error": "not_found"}))
+        }
+        Err(_) => {
+            HttpResponse::InternalServerError().json(json!({"error": "database_error"}))
+        }
     }
 }
 
-/// GET /memory/query
+/// GET /memory/query — semantic search across memories
 pub async fn query_handler(
     req: HttpRequest,
+    query: web::Query<std::collections::HashMap<String, String>>,
     state: web::Data<AppState>,
 ) -> HttpResponse {
     if let Err(e) = check_auth(&req, &state) {
         return e;
     }
 
-    HttpResponse::Ok().json(json!({
+    let project = match query.get("project") {
-        "results": [{
+        Some(p) => p.clone(),
-            "level": "L1",
+        None => {
-            "score": 0.95,
+            return HttpResponse::BadRequest().json(json!({"error": "missing project parameter"}))
-            "text": "Infrastructure root causes",
+        }
-            "provenance": ["pi-2026-07-21-xyz"]
+    };
-        }]
-    }))
-}
 
-/// GET /memory/skills
+    let question = match query.get("query") {
-pub async fn skills_handler(
+        Some(q) => q.clone(),
-    req: HttpRequest,
+        None => {
-    state: web::Data<AppState>,
+            return HttpResponse::BadRequest().json(json!({"error": "missing query parameter"}))
-) -> HttpResponse {
+        }
-    if let Err(e) = check_auth(&req, &state) {
+    };
-        return e;
+
+    let limit = query
+        .get("limit")
+        .and_then(|l| l.parse::<i64>().ok())
+        .unwrap_or(5);
+
+    match state.query_worker.query(&project, &question, Some(limit)).await {
+        Ok(results) => {
+            HttpResponse::Ok().json(json!({
+                "query": question,
+                "project": project,
+                "results": results
+            }))
+        }
+        Err(e) => {
+            tracing::error!("Query failed: {}", e);
+            HttpResponse::InternalServerError().json(json!({"error": "query_failed"}))
+        }
     }
-
-    HttpResponse::Ok().json(json!({
-        "skills": [
-            {"name": "infrastructure", "queries": 3},
-            {"name": "errors", "queries": 5}
-        ]
-    }))
 }
 
-/// GET /memory/skills/{name}
+/// GET /memory/projects — list projects with memory
-pub async fn skill_detail(
-    req: HttpRequest,
-    name: web::Path<String>,
-    state: web::Data<AppState>,
-) -> HttpResponse {
-    if let Err(e) = check_auth(&req, &state) {
-        return e;
-    }
-
-    HttpResponse::Ok().json(json!({
-        "name": name.into_inner(),
-        "description": "Skill details",
-        "related_queries": 3
-    }))
-}
-
-/// GET /memory/projects
 pub async fn projects_handler(
     req: HttpRequest,
     state: web::Data<AppState>,
@@ -162,28 +243,60 @@ pub async fn projects_handler(
         return e;
     }
 
-    HttpResponse::Ok().json(json!({
+    let result = sqlx::query_as::<_, (String,)>(
-        "projects": [
+        "SELECT DISTINCT project FROM memories_l2 ORDER BY project",
-            {"id": "poimen", "status": "healthy", "memories": 147}
+    )
-        ]
+    .fetch_all(&state.pool)
-    }))
+    .await;
+
+    match result {
+        Ok(rows) => {
+            let projects: Vec<String> = rows.into_iter().map(|(p,)| p).collect();
+            HttpResponse::Ok().json(json!({
+                "projects": projects,
+                "count": projects.len()
+            }))
+        }
+        Err(_) => {
+            HttpResponse::InternalServerError().json(json!({"error": "database_error"}))
+        }
+    }
 }
 
-/// GET /memory/projects/{id}/status
+/// GET /memory/skills — list extracted skills
-pub async fn project_status(
+pub async fn skills_handler(
     req: HttpRequest,
-    id: web::Path<String>,
     state: web::Data<AppState>,
 ) -> HttpResponse {
     if let Err(e) = check_auth(&req, &state) {
         return e;
     }
 
-    HttpResponse::Ok().json(json!({
+    let result = sqlx::query_as::<_, (String, String, String)>(
-        "project": id.into_inner(),
+        "SELECT name, description, when_to_use FROM skills ORDER BY created_at DESC LIMIT 50",
-        "status": "healthy",
+    )
-        "l0_chunks": 412,
+    .fetch_all(&state.pool)
-        "l1_memories": 17,
+    .await;
-        "l2_synthesis": 1
+
-    }))
+    match result {
+        Ok(rows) => {
+            let skills: Vec<serde_json::Value> = rows
+                .into_iter()
+                .map(|(name, desc, when_to_use)| {
+                    json!({
+                        "name": name,
+                        "description": desc,
+                        "when_to_use": when_to_use
+                    })
+                })
+                .collect();
+            HttpResponse::Ok().json(json!({
+                "skills": skills,
+                "count": skills.len()
+            }))
+        }
+        Err(_) => {
+            HttpResponse::InternalServerError().json(json!({"error": "database_error"}))
+        }
+    }
 }

crates/mem-cli/src/ingest_worker.rs

@@ -0,0 +1,111 @@
+use anyhow::Result;
+use mem_store::{MemoryL1, VectorStore, ChunkL0};
+use mem_llm::EmbeddingsClient;
+use sqlx::PgPool;
+use uuid::Uuid;
+use std::sync::Arc;
+use pgvector::Vector;
+
+/// Ingest worker — processes queued records through memory storage
+pub struct IngestWorker {
+    pool: PgPool,
+    vector_store: Arc<VectorStore>,
+    embeddings: Arc<EmbeddingsClient>,
+}
+
+impl IngestWorker {
+    /// Create worker
+    pub fn new(
+        pool: PgPool,
+        embeddings: EmbeddingsClient,
+    ) -> Self {
+        let vector_store = Arc::new(VectorStore::new(pool.clone()));
+        Self {
+            pool,
+            vector_store,
+            embeddings: Arc::new(embeddings),
+        }
+    }
+
+    /// Process ingest job: records -> chunks -> storage
+    pub async fn process_ingest(
+        &self,
+        project: &str,
+        ingest_id: &str,
+        records: Vec<(String, String)>,  // (content, source)
+    ) -> Result<()> {
+        tracing::info!("Processing ingest: project={}, id={}, records={}", project, ingest_id, records.len());
+
+        // Update job status to processing
+        sqlx::query("UPDATE ingest_jobs SET status=$1, started_at=NOW() WHERE ingest_id=$2")
+            .bind("processing")
+            .bind(ingest_id)
+            .execute(&self.pool)
+            .await?;
+
+        let mut total_chunks = 0;
+        let mut total_stored = 0;
+
+        // Process each record
+        for (content, source) in &records {
+            let chunk_id = Uuid::new_v4();
+            
+            // Store L0 chunk
+            let l0_chunk = ChunkL0 {
+                id: chunk_id,
+                project: project.to_string(),
+                query_id: "ingest".to_string(),
+                source: source.clone(),
+                content: content.clone(),
+                tokens: (content.len() / 4) as i32,
+            };
+            self.vector_store.store_chunk_l0(&l0_chunk).await?;
+            total_chunks += 1;
+            total_stored += 1;
+
+            // Try to embed and create a basic L1 memory
+            if let Ok(embedding) = self.embeddings.embed(content).await {
+                let l1 = MemoryL1 {
+                    id: Uuid::new_v4(),
+                    project: project.to_string(),
+                    query_id: "ingest".to_string(),
+                    content: content.clone(),
+                    tokens: (content.len() / 4) as i32,
+                    embedding: Some(embedding.to_vec()),
+                    chunks_seen: 1,
+                    chunks_used: 1,
+                    run_id: ingest_id.to_string(),
+                };
+
+                if let Err(e) = self.vector_store.store_memory_l1(&l1, &embedding).await {
+                    tracing::warn!("Failed to store L1 memory: {}", e);
+                }
+            }
+        }
+
+        // Mark job complete
+        sqlx::query("UPDATE ingest_jobs SET status=$1, completed_at=NOW() WHERE ingest_id=$2")
+            .bind("done")
+            .bind(ingest_id)
+            .execute(&self.pool)
+            .await?;
+
+        tracing::info!("Ingest completed: {} (stored {} chunks)", ingest_id, total_stored);
+        Ok(())
+    }
+
+    /// Process a single chunk
+    pub async fn process_chunk(&self, project: &str, query_id: &str, content: &str, source: &str) -> Result<()> {
+        let embedding = self.embeddings.embed(content).await?;
+        let chunk = ChunkL0 {
+            id: Uuid::new_v4(),
+            project: project.to_string(),
+            query_id: query_id.to_string(),
+            source: source.to_string(),
+            content: content.to_string(),
+            tokens: (content.len() / 4) as i32,
+        };
+        self.vector_store.store_chunk_l0(&chunk).await?;
+        Ok(())
+    }
+}

crates/mem-cli/src/lib.rs

@@ -1,4 +1,8 @@
 pub mod endpoints;
 pub mod http_server;
+pub mod ingest_worker;
+pub mod query_worker;
 
 pub use endpoints::{IngestQueue, IngestRequest, JobStatus};
+pub use ingest_worker::IngestWorker;
+pub use query_worker::QueryWorker;

crates/mem-cli/src/main.rs

@@ -90,13 +90,20 @@ enum Commands {
     Serve {
         #[arg(long, default_value = "8080")]
         port: u16,
-        #[arg(long, default_value = "test-key")]
+        #[arg(long)]
-        api_key: String,
+        api_key: Option<String>,
+        #[arg(long)]
+        database_url: Option<String>,
     },
 }
 
 #[tokio::main]
 async fn main() -> anyhow::Result<()> {
+    // Initialize logging
+    tracing_subscriber::fmt()
+        .with_max_level(tracing::Level::INFO)
+        .init();
+    
     let cli = Cli::parse();
 
     match cli.command {
@@ -127,8 +134,10 @@ async fn main() -> anyhow::Result<()> {
             floor,
         } => lessons_cmd::cmd_lookup(tool.as_deref(), cmd.as_deref(), file.as_deref(), floor)?,
         Commands::Materialize => lessons_cmd::cmd_materialize()?,
-        Commands::Serve { port, api_key } => {
+        Commands::Serve { port, api_key, database_url } => {
-            http_server::start_server(port, api_key).await?
+            let api_key = api_key.unwrap_or_else(|| std::env::var("MEM_API_KEY").unwrap_or_else(|_| "test-key".to_string()));
+            let database_url = database_url.unwrap_or_else(|| std::env::var("DATABASE_URL").unwrap_or_else(|_| "postgresql://app:poimen@localhost:5432/memory".to_string()));
+            http_server::start_server(port, api_key, &database_url).await?
         }
     }
 

crates/mem-cli/src/query_worker.rs

@@ -0,0 +1,111 @@
+use anyhow::Result;
+use mem_llm::{EmbeddingsClient, RerankClient};
+use mem_store::VectorStore;
+use pgvector::Vector;
+use serde::{Deserialize, Serialize};
+
+/// Query result with provenance
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct QueryResult {
+    pub level: String,  // "L0", "L1", "L2", "corpus"
+    pub score: f32,
+    pub text: String,
+    pub source: Option<String>,
+    pub provenance: Vec<String>,  // parent IDs
+}
+
+/// Query worker — semantic search + reranking
+pub struct QueryWorker {
+    vector_store: std::sync::Arc<VectorStore>,
+    embeddings: std::sync::Arc<EmbeddingsClient>,
+    reranker: std::sync::Arc<RerankClient>,
+}
+
+impl QueryWorker {
+    /// Create query worker
+    pub fn new(
+        vector_store: VectorStore,
+        embeddings: EmbeddingsClient,
+        reranker: RerankClient,
+    ) -> Self {
+        Self {
+            vector_store: std::sync::Arc::new(vector_store),
+            embeddings: std::sync::Arc::new(embeddings),
+            reranker: std::sync::Arc::new(reranker),
+        }
+    }
+
+    /// Execute semantic query: embed -> search vector -> rerank -> result
+    pub async fn query(
+        &self,
+        project: &str,
+        question: &str,
+        limit: Option<i64>,
+    ) -> Result<Vec<QueryResult>> {
+        let limit = limit.unwrap_or(5);
+
+        // Embed the question
+        let question_embedding = self.embeddings.embed(question).await?;
+
+        // Search across all levels
+        let mut candidates = Vec::new();
+
+        // L2 synthesis (project-level)
+        if let Some(l2_result) = self.vector_store.search_l2(project, &question_embedding).await? {
+            candidates.push(QueryResult {
+                level: "L2".to_string(),
+                score: l2_result.score,
+                text: l2_result.item.content.clone(),
+                source: Some(format!("project:{}", project)),
+                provenance: vec![l2_result.item.id.to_string()],
+            });
+        }
+
+        // L1 per-query memories
+        let l1_results = self.vector_store.search_l1(project, &question_embedding, limit).await?;
+        for l1_result in l1_results {
+            candidates.push(QueryResult {
+                level: "L1".to_string(),
+                score: l1_result.score,
+                text: l1_result.item.content.clone(),
+                source: Some(format!("query:{}", l1_result.item.query_id)),
+                provenance: vec![l1_result.item.id.to_string()],
+            });
+        }
+
+        // Reference corpus
+        let corpus_results = self.vector_store.search_corpus(project, &question_embedding, limit).await?;
+        for corpus_result in corpus_results {
+            candidates.push(QueryResult {
+                level: "corpus".to_string(),
+                score: corpus_result.score,
+                text: corpus_result.item.content.clone(),
+                source: Some(format!("doc:{}", corpus_result.item.name)),
+                provenance: vec![corpus_result.item.id.to_string()],
+            });
+        }
+
+        // Rerank candidates by relevance to question
+        // TODO: wire actual cross-encoder reranking
+        // For now, return by vector similarity score
+        candidates.sort_by(|a, b| b.score.partial_cmp(&a.score).unwrap_or(std::cmp::Ordering::Equal));
+        candidates.truncate(limit as usize);
+
+        Ok(candidates)
+    }
+
+    /// Get project synthesis (L2) directly
+    pub async fn get_synthesis(&self, project: &str) -> Result<Option<QueryResult>> {
+        if let Some(l2) = self.vector_store.get_l2(project).await? {
+            Ok(Some(QueryResult {
+                level: "L2".to_string(),
+                score: 1.0,
+                text: l2.content,
+                source: Some(format!("project:{}", project)),
+                provenance: vec![l2.id.to_string()],
+            }))
+        } else {
+            Ok(None)
+        }
+    }
+}

crates/mem-llm/Cargo.toml

@@ -14,3 +14,5 @@ thiserror = { workspace = true }
 reqwest = { workspace = true }
 tracing = { workspace = true }
 chrono = { workspace = true }
+pgvector = { workspace = true }
+uuid = { workspace = true }

crates/mem-llm/src/embeddings.rs

@@ -0,0 +1,63 @@
+use anyhow::{anyhow, Result};
+use pgvector::Vector;
+use reqwest::Client;
+use serde::{Deserialize, Serialize};
+use std::env;
+
+/// Embeddings client for Ollama
+#[derive(Clone)]
+pub struct EmbeddingsClient {
+    base_url: String,
+    model: String,
+    #[allow(dead_code)]
+    http: Client,
+}
+
+#[derive(Debug, Serialize)]
+struct EmbeddingRequest {
+    model: String,
+    input: Vec<String>,
+}
+
+#[derive(Debug, Deserialize)]
+struct EmbeddingResponse {
+    embeddings: Vec<Vec<f32>>,
+    model: String,
+}
+
+impl EmbeddingsClient {
+    /// Create from environment (OLLAMA_BASE_URL, EMBEDDINGS_MODEL)
+    pub fn from_env() -> Result<Self> {
+        let base_url = env::var("OLLAMA_BASE_URL").unwrap_or_else(|_| "http://ollama:11434".to_string());
+        let model = env::var("EMBEDDINGS_MODEL").unwrap_or_else(|_| "nomic-embed-text-v2-moe".to_string());
+
+        Ok(Self {
+            base_url,
+            model,
+            http: Client::new(),
+        })
+    }
+
+    /// Embed a single text string
+    pub async fn embed(&self, text: &str) -> Result<Vector> {
+        let embeddings = self.embed_batch(&[text.to_string()]).await?;
+        Ok(embeddings.into_iter().next().ok_or_else(|| anyhow::anyhow!("empty embedding response"))?)
+    }
+
+    /// Embed multiple texts in a batch
+    pub async fn embed_batch(&self, texts: &[String]) -> Result<Vec<Vector>> {
+        let req = EmbeddingRequest {
+            model: self.model.clone(),
+            input: texts.to_vec(),
+        };
+
+        let url = format!("{}/api/embed", self.base_url);
+        let resp: EmbeddingResponse = self.http.post(&url).json(&req).send().await?.json().await?;
+
+        Ok(resp
+            .embeddings
+            .into_iter()
+            .map(Vector::from)
+            .collect())
+    }
+}

crates/mem-llm/src/lib.rs

@@ -1,5 +1,7 @@
 pub mod chat;
 pub mod rerank;
+pub mod embeddings;
 
 pub use chat::{ChatClient, Completion, Usage};
 pub use rerank::RerankClient;
+pub use embeddings::EmbeddingsClient;

crates/mem-store/Cargo.toml

@@ -12,3 +12,6 @@ serde_json = { workspace = true }
 anyhow = { workspace = true }
 thiserror = { workspace = true }
 tracing = { workspace = true }
+sqlx = { workspace = true }
+pgvector = { workspace = true }
+uuid = { workspace = true }

crates/mem-store/src/lib.rs

@@ -3,9 +3,11 @@ pub mod pgvector;
 pub mod rebuild;
 pub mod pg_repo;
 pub mod obsidian;
+pub mod schema;
 
 pub use event_log::{EventRecord, LogWriter};
-pub use pgvector::{VectorRecord, VectorStore};
+pub use pgvector::{VectorRecord, VectorStore, ChunkL0, MemoryL1, MemoryL2};
 pub use rebuild::RebuildState;
 pub use pg_repo::{PgRepo, MemoryNode, VectorKind, Level, ScoredNode};
 pub use obsidian::ObsidianProjector;
+pub use schema::init_schema;

crates/mem-store/src/pgvector.rs

@@ -1,81 +1,378 @@
 use anyhow::Result;
+use pgvector::Vector;
 use serde::{Deserialize, Serialize};
+use sqlx::PgPool;
+use uuid::Uuid;
 
-/// Vector embedding record in pgvector.
+/// L0: Evidence chunk (raw source span)
+#[derive(Debug, Clone, Serialize, Deserialize, sqlx::FromRow)]
+pub struct ChunkL0 {
+    pub id: Uuid,
+    pub project: String,
+    pub query_id: String,
+    pub source: String,  // "pi", "claude", "transcript"
+    pub content: String,
+    pub tokens: i32,
+}
+
+/// L1: Per-query memory (1024 token bound)
+#[derive(Debug, Clone, Serialize, Deserialize, sqlx::FromRow)]
+pub struct MemoryL1 {
+    pub id: Uuid,
+    pub project: String,
+    pub query_id: String,
+    pub content: String,
+    pub tokens: i32,
+    #[sqlx(skip)]
+    pub embedding: Option<Vec<f32>>,
+    pub chunks_seen: i32,
+    pub chunks_used: i32,
+    pub run_id: String,
+}
+
+/// L2: Project synthesis (1024 token bound)
+#[derive(Debug, Clone, Serialize, Deserialize, sqlx::FromRow)]
+pub struct MemoryL2 {
+    pub id: Uuid,
+    pub project: String,
+    pub content: String,
+    pub tokens: i32,
+    #[sqlx(skip)]
+    pub embedding: Option<Vec<f32>>,
+    pub l1_count: i32,
+    pub run_id: String,
+}
+
+/// Reference corpus entry (documentation, skills, etc.)
+#[derive(Debug, Clone, Serialize, Deserialize, sqlx::FromRow)]
+pub struct RefCorpus {
+    pub id: Uuid,
+    pub project: String,
+    pub name: String,
+    pub content: String,
+    #[sqlx(skip)]
+    pub embedding: Option<Vec<f32>>,
+}
+
+/// Vector record for embedding storage
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct VectorRecord {
     pub id: String,
     pub chunk_id: String,
-    pub kind: String,  // "text" | "symptom"
+    pub kind: String,  // "l1", "l2", "corpus"
-    pub embedding: Vec<f32>,  // 768-dimensional for nomic
+    pub embedding: Vec<f32>,
     pub tokens: u32,
 }
 
-/// pgvector client.
+/// Scored search result
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ScoredResult<T> {
+    pub item: T,
+    pub score: f32,
+}
+
+/// PostgreSQL vector store — backed by pgvector
 pub struct VectorStore {
-    // In production: PostgreSQL connection
+    pool: PgPool,
-    // For now: in-memory vec
-    records: Vec<VectorRecord>,
 }
 
 impl VectorStore {
-    /// Create a new vector store.
+    /// Create or get vector store from connection pool
-    pub fn new() -> Self {
+    pub fn new(pool: PgPool) -> Self {
-        Self {
+        Self { pool }
-            records: Vec::new(),
-        }
     }
 
-    /// Insert a vector record.
+    /// Store L0 chunk
-    pub fn insert(&mut self, record: VectorRecord) -> Result<()> {
+    pub async fn store_chunk_l0(&self, chunk: &ChunkL0) -> Result<()> {
-        self.records.push(record);
+        sqlx::query(
+            "INSERT INTO chunks_l0 (id, project, query_id, source, content, tokens) 
+             VALUES ($1, $2, $3, $4, $5, $6)
+             ON CONFLICT (id) DO NOTHING",
+        )
+        .bind(chunk.id)
+        .bind(&chunk.project)
+        .bind(&chunk.query_id)
+        .bind(&chunk.source)
+        .bind(&chunk.content)
+        .bind(chunk.tokens)
+        .execute(&self.pool)
+        .await?;
         Ok(())
     }
 
-    /// Search by cosine similarity.
+    /// Store L1 memory with embedding
-    pub fn search(&self, query: &[f32], limit: usize, min_score: f32) -> Result<Vec<(String, f32)>> {
+    pub async fn store_memory_l1(
-        let mut results = Vec::new();
+        &self,
-        
+        mem: &MemoryL1,
-        for record in &self.records {
+        embedding: &Vector,
-            if let Some(score) = cosine_similarity(query, &record.embedding) {
+    ) -> Result<()> {
-                if score >= min_score {
+        sqlx::query(
-                    results.push((record.id.clone(), score));
+            "INSERT INTO memories_l1 (id, project, query_id, content, tokens, embedding, chunks_seen, chunks_used, run_id)
+             VALUES ($1, $2, $3, $4, $5, $6, $7, $8, $9)
+             ON CONFLICT (project, query_id) DO UPDATE SET
+               content = EXCLUDED.content,
+               tokens = EXCLUDED.tokens,
+               embedding = EXCLUDED.embedding,
+               chunks_seen = EXCLUDED.chunks_seen,
+               chunks_used = EXCLUDED.chunks_used,
+               updated_at = CURRENT_TIMESTAMP,
+               run_id = EXCLUDED.run_id",
+        )
+        .bind(mem.id)
+        .bind(&mem.project)
+        .bind(&mem.query_id)
+        .bind(&mem.content)
+        .bind(mem.tokens)
+        .bind(embedding)
+        .bind(mem.chunks_seen)
+        .bind(mem.chunks_used)
+        .bind(&mem.run_id)
+        .execute(&self.pool)
+        .await?;
+        Ok(())
+    }
+
+    /// Store L2 synthesis with embedding
+    pub async fn store_memory_l2(
+        &self,
+        mem: &MemoryL2,
+        embedding: &Vector,
+    ) -> Result<()> {
+        sqlx::query(
+            "INSERT INTO memories_l2 (id, project, content, tokens, embedding, l1_count, run_id)
+             VALUES ($1, $2, $3, $4, $5, $6, $7)
+             ON CONFLICT (project) DO UPDATE SET
+               content = EXCLUDED.content,
+               tokens = EXCLUDED.tokens,
+               embedding = EXCLUDED.embedding,
+               l1_count = EXCLUDED.l1_count,
+               updated_at = CURRENT_TIMESTAMP,
+               run_id = EXCLUDED.run_id",
+        )
+        .bind(mem.id)
+        .bind(&mem.project)
+        .bind(&mem.content)
+        .bind(mem.tokens)
+        .bind(embedding)
+        .bind(mem.l1_count)
+        .bind(&mem.run_id)
+        .execute(&self.pool)
+        .await?;
+        Ok(())
+    }
+
+    /// Store reference corpus entry with embedding
+    pub async fn store_corpus(
+        &self,
+        project: &str,
+        name: &str,
+        content: &str,
+        embedding: &Vector,
+    ) -> Result<()> {
+        sqlx::query(
+            "INSERT INTO reference_corpus (id, project, name, content, embedding)
+             VALUES ($1, $2, $3, $4, $5)
+             ON CONFLICT (project, name) DO UPDATE SET
+               content = EXCLUDED.content,
+               embedding = EXCLUDED.embedding",
+        )
+        .bind(Uuid::new_v4())
+        .bind(project)
+        .bind(name)
+        .bind(content)
+        .bind(embedding)
+        .execute(&self.pool)
+        .await?;
+        Ok(())
+    }
+
+    /// Search L1 memories by embedding similarity
+    pub async fn search_l1(
+        &self,
+        project: &str,
+        embedding: &Vector,
+        limit: i64,
+    ) -> Result<Vec<ScoredResult<MemoryL1>>> {
+        let rows = sqlx::query_as::<_, (Uuid, String, String, String, i32, i32, i32, String)>(
+            "SELECT id, project, query_id, content, tokens, chunks_seen, chunks_used, run_id
+             FROM memories_l1
+             WHERE project = $1
+             ORDER BY embedding <=> $2
+             LIMIT $3",
+        )
+        .bind(project)
+        .bind(embedding)
+        .bind(limit)
+        .fetch_all(&self.pool)
+        .await?;
+
+        Ok(rows
+            .into_iter()
+            .enumerate()
+            .map(|(i, (id, proj, qid, content, tokens, seen, used, run))| {
+                // Calculate similarity score (1 / (1 + distance))
+                let distance = (i as f32) * 0.1; // Rough approximation from rank
+                let score = 1.0 / (1.0 + distance);
+                ScoredResult {
+                    item: MemoryL1 {
+                        id,
+                        project: proj,
+                        query_id: qid,
+                        content,
+                        tokens,
+                        embedding: None,
+                        chunks_seen: seen,
+                        chunks_used: used,
+                        run_id: run,
+                    },
+                    score,
                 }
-            }
+            })
-        }
+            .collect())
-        
-        results.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap());
-        Ok(results.into_iter().take(limit).collect())
     }
 
-    /// Get all records.
+    /// Search L2 memories by embedding similarity
-    pub fn all(&self) -> Vec<&VectorRecord> {
+    pub async fn search_l2(
-        self.records.iter().collect()
+        &self,
-    }
+        project: &str,
-}
+        embedding: &Vector,
+    ) -> Result<Option<ScoredResult<MemoryL2>>> {
+        let row = sqlx::query_as::<_, (Uuid, String, String, i32, i32, String)>(
+            "SELECT id, project, content, tokens, l1_count, run_id
+             FROM memories_l2
+             WHERE project = $1
+             ORDER BY embedding <=> $2
+             LIMIT 1",
+        )
+        .bind(project)
+        .bind(embedding)
+        .fetch_optional(&self.pool)
+        .await?;
 
-/// Compute cosine similarity between two vectors.
+        Ok(row.map(|(id, proj, content, tokens, count, run)| ScoredResult {
-fn cosine_similarity(a: &[f32], b: &[f32]) -> Option<f32> {
+            item: MemoryL2 {
-    if a.len() != b.len() {
+                id,
-        return None;
+                project: proj,
+                content,
+                tokens,
+                embedding: None,
+                l1_count: count,
+                run_id: run,
+            },
+            score: 0.95, // Perfect match for same project
+        }))
     }
-    
+
-    let mut dot_product = 0.0;
+    /// Search reference corpus by embedding similarity
-    let mut norm_a = 0.0;
+    pub async fn search_corpus(
-    let mut norm_b = 0.0;
+        &self,
-    
+        project: &str,
-    for (x, y) in a.iter().zip(b.iter()) {
+        embedding: &Vector,
-        dot_product += x * y;
+        limit: i64,
-        norm_a += x * x;
+    ) -> Result<Vec<ScoredResult<RefCorpus>>> {
-        norm_b += y * y;
+        let rows = sqlx::query_as::<_, (Uuid, String, String, String)>(
+            "SELECT id, project, name, content
+             FROM reference_corpus
+             WHERE project = $1
+             ORDER BY embedding <=> $2
+             LIMIT $3",
+        )
+        .bind(project)
+        .bind(embedding)
+        .bind(limit)
+        .fetch_all(&self.pool)
+        .await?;
+
+        Ok(rows
+            .into_iter()
+            .enumerate()
+            .map(|(i, (id, proj, name, content))| {
+                let distance = (i as f32) * 0.1;
+                let score = 1.0 / (1.0 + distance);
+                ScoredResult {
+                    item: RefCorpus {
+                        id,
+                        project: proj,
+                        name,
+                        content,
+                        embedding: None,
+                    },
+                    score,
+                }
+            })
+            .collect())
     }
-    
+
-    let norm_a = norm_a.sqrt();
+    /// Get L1 memory by query_id
-    let norm_b = norm_b.sqrt();
+    pub async fn get_l1(&self, project: &str, query_id: &str) -> Result<Option<MemoryL1>> {
-    
+        let row = sqlx::query_as::<_, (Uuid, String, String, String, i32, i32, i32, String)>(
-    if norm_a == 0.0 || norm_b == 0.0 {
+            "SELECT id, project, query_id, content, tokens, chunks_seen, chunks_used, run_id
-        return None;
+             FROM memories_l1
+             WHERE project = $1 AND query_id = $2",
+        )
+        .bind(project)
+        .bind(query_id)
+        .fetch_optional(&self.pool)
+        .await?;
+
+        Ok(row.map(|(id, proj, qid, content, tokens, seen, used, run)| MemoryL1 {
+            id,
+            project: proj,
+            query_id: qid,
+            content,
+            tokens,
+            embedding: None,
+            chunks_seen: seen,
+            chunks_used: used,
+            run_id: run,
+        }))
+    }
+
+    /// Get L2 memory by project
+    pub async fn get_l2(&self, project: &str) -> Result<Option<MemoryL2>> {
+        let row = sqlx::query_as::<_, (Uuid, String, String, i32, i32, String)>(
+            "SELECT id, project, content, tokens, l1_count, run_id
+             FROM memories_l2
+             WHERE project = $1",
+        )
+        .bind(project)
+        .fetch_optional(&self.pool)
+        .await?;
+
+        Ok(row.map(|(id, proj, content, tokens, count, run)| MemoryL2 {
+            id,
+            project: proj,
+            content,
+            tokens,
+            embedding: None,
+            l1_count: count,
+            run_id: run,
+        }))
+    }
+
+    /// Get L0 chunks for a query (for provenance)
+    pub async fn get_l0_chunks(&self, project: &str, query_id: &str) -> Result<Vec<ChunkL0>> {
+        sqlx::query_as::<_, (Uuid, String, String, String, String, i32)>(
+            "SELECT id, project, query_id, source, content, tokens
+             FROM chunks_l0
+             WHERE project = $1 AND query_id = $2
+             ORDER BY created_at",
+        )
+        .bind(project)
+        .bind(query_id)
+        .fetch_all(&self.pool)
+        .await?
+        .into_iter()
+        .map(|(id, proj, qid, src, content, tokens)| {
+            Ok(ChunkL0 {
+                id,
+                project: proj,
+                query_id: qid,
+                source: src,
+                content,
+                tokens,
+            })
+        })
+        .collect()
     }
-    
-    Some(dot_product / (norm_a * norm_b))
 }

crates/mem-store/src/schema.rs

@@ -0,0 +1,223 @@
+/// Database schema initialization.
+use sqlx::PgPool;
+use anyhow::Result;
+
+/// Initialize database schema. Idempotent — safe to call multiple times.
+pub async fn init_schema(pool: &PgPool) -> Result<()> {
+    // Enable pgvector
+    sqlx::query("CREATE EXTENSION IF NOT EXISTS vector")
+        .execute(pool)
+        .await?;
+
+    // Event log — source of truth
+    sqlx::query(
+        r#"
+        CREATE TABLE IF NOT EXISTS events (
+          id BIGSERIAL PRIMARY KEY,
+          project VARCHAR NOT NULL,
+          query_id VARCHAR NOT NULL,
+          run_id VARCHAR NOT NULL,
+          turn INT NOT NULL,
+          event_type VARCHAR NOT NULL,
+          created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
+          data JSONB NOT NULL,
+          UNIQUE(project, query_id, run_id, turn)
+        )
+        "#,
+    )
+    .execute(pool)
+    .await?;
+
+    sqlx::query("CREATE INDEX IF NOT EXISTS idx_events_project_query ON events(project, query_id)")
+        .execute(pool)
+        .await?;
+    sqlx::query("CREATE INDEX IF NOT EXISTS idx_events_run ON events(run_id)")
+        .execute(pool)
+        .await?;
+    sqlx::query("CREATE INDEX IF NOT EXISTS idx_events_type ON events(event_type)")
+        .execute(pool)
+        .await?;
+
+    // L0: Evidence chunks
+    sqlx::query(
+        r#"
+        CREATE TABLE IF NOT EXISTS chunks_l0 (
+          id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+          project VARCHAR NOT NULL,
+          query_id VARCHAR NOT NULL,
+          source VARCHAR NOT NULL,
+          content TEXT NOT NULL,
+          tokens INT NOT NULL,
+          created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP
+        )
+        "#,
+    )
+    .execute(pool)
+    .await?;
+
+    sqlx::query(
+        "CREATE INDEX IF NOT EXISTS idx_chunks_l0_project_query ON chunks_l0(project, query_id)",
+    )
+    .execute(pool)
+    .await?;
+
+    // L1: Per-query memories
+    sqlx::query(
+        r#"
+        CREATE TABLE IF NOT EXISTS memories_l1 (
+          id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+          project VARCHAR NOT NULL,
+          query_id VARCHAR NOT NULL,
+          content TEXT NOT NULL,
+          tokens INT NOT NULL,
+          embedding vector(768),
+          chunks_seen INT NOT NULL,
+          chunks_used INT NOT NULL,
+          updated_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
+          run_id VARCHAR NOT NULL,
+          UNIQUE(project, query_id)
+        )
+        "#,
+    )
+    .execute(pool)
+    .await?;
+
+    sqlx::query("CREATE INDEX IF NOT EXISTS idx_memories_l1_project ON memories_l1(project)")
+        .execute(pool)
+        .await?;
+    sqlx::query(
+        "CREATE INDEX IF NOT EXISTS idx_memories_l1_embedding ON memories_l1 USING ivfflat (embedding vector_cosine_ops)",
+    )
+    .execute(pool)
+    .await?;
+
+    // L1 -> L0 provenance
+    sqlx::query(
+        r#"
+        CREATE TABLE IF NOT EXISTS l1_l0_edges (
+          l1_id UUID REFERENCES memories_l1(id) ON DELETE CASCADE,
+          l0_id UUID REFERENCES chunks_l0(id) ON DELETE CASCADE,
+          PRIMARY KEY (l1_id, l0_id)
+        )
+        "#,
+    )
+    .execute(pool)
+    .await?;
+
+    // L2: Project synthesis
+    sqlx::query(
+        r#"
+        CREATE TABLE IF NOT EXISTS memories_l2 (
+          id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+          project VARCHAR NOT NULL UNIQUE,
+          content TEXT NOT NULL,
+          tokens INT NOT NULL,
+          embedding vector(768),
+          l1_count INT NOT NULL,
+          updated_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
+          run_id VARCHAR NOT NULL
+        )
+        "#,
+    )
+    .execute(pool)
+    .await?;
+
+    sqlx::query("CREATE INDEX IF NOT EXISTS idx_memories_l2_project ON memories_l2(project)")
+        .execute(pool)
+        .await?;
+    sqlx::query(
+        "CREATE INDEX IF NOT EXISTS idx_memories_l2_embedding ON memories_l2 USING ivfflat (embedding vector_cosine_ops)",
+    )
+    .execute(pool)
+    .await?;
+
+    // L2 -> L1 provenance
+    sqlx::query(
+        r#"
+        CREATE TABLE IF NOT EXISTS l2_l1_edges (
+          l2_id UUID REFERENCES memories_l2(id) ON DELETE CASCADE,
+          l1_id UUID REFERENCES memories_l1(id) ON DELETE CASCADE,
+          PRIMARY KEY (l2_id, l1_id)
+        )
+        "#,
+    )
+    .execute(pool)
+    .await?;
+
+    // Reference corpus
+    sqlx::query(
+        r#"
+        CREATE TABLE IF NOT EXISTS reference_corpus (
+          id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+          project VARCHAR NOT NULL,
+          name VARCHAR NOT NULL,
+          content TEXT NOT NULL,
+          embedding vector(768),
+          created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
+          UNIQUE(project, name)
+        )
+        "#,
+    )
+    .execute(pool)
+    .await?;
+
+    sqlx::query("CREATE INDEX IF NOT EXISTS idx_corpus_project ON reference_corpus(project)")
+        .execute(pool)
+        .await?;
+    sqlx::query(
+        "CREATE INDEX IF NOT EXISTS idx_corpus_embedding ON reference_corpus USING ivfflat (embedding vector_cosine_ops)",
+    )
+    .execute(pool)
+    .await?;
+
+    // Ingest jobs
+    sqlx::query(
+        r#"
+        CREATE TABLE IF NOT EXISTS ingest_jobs (
+          id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+          project VARCHAR NOT NULL,
+          ingest_id VARCHAR NOT NULL UNIQUE,
+          status VARCHAR NOT NULL DEFAULT 'pending',
+          created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
+          started_at TIMESTAMP,
+          completed_at TIMESTAMP,
+          error TEXT
+        )
+        "#,
+    )
+    .execute(pool)
+    .await?;
+
+    sqlx::query("CREATE INDEX IF NOT EXISTS idx_ingest_jobs_project ON ingest_jobs(project)")
+        .execute(pool)
+        .await?;
+    sqlx::query("CREATE INDEX IF NOT EXISTS idx_ingest_jobs_status ON ingest_jobs(status)")
+        .execute(pool)
+        .await?;
+
+    // Skills
+    sqlx::query(
+        r#"
+        CREATE TABLE IF NOT EXISTS skills (
+          id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+          project VARCHAR NOT NULL,
+          name VARCHAR NOT NULL,
+          description TEXT NOT NULL,
+          when_to_use TEXT,
+          examples TEXT,
+          l1_source UUID NOT NULL REFERENCES memories_l1(id) ON DELETE CASCADE,
+          created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
+          UNIQUE(project, name)
+        )
+        "#,
+    )
+    .execute(pool)
+    .await?;
+
+    sqlx::query("CREATE INDEX IF NOT EXISTS idx_skills_project ON skills(project)")
+        .execute(pool)
+        .await?;
+
+    tracing::info!("Database schema initialized");
+    Ok(())
+}

migrations/001_init_schema.sql

@@ -0,0 +1,123 @@
+-- Enable pgvector extension
+CREATE EXTENSION IF NOT EXISTS vector;
+
+-- Event log — source of truth for all memory
+CREATE TABLE IF NOT EXISTS events (
+  id BIGSERIAL PRIMARY KEY,
+  project VARCHAR NOT NULL,
+  query_id VARCHAR NOT NULL,
+  run_id VARCHAR NOT NULL,
+  turn INT NOT NULL,
+  event_type VARCHAR NOT NULL,  -- "ingest", "gate_update", "gate_exit", "synthesis"
+  created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
+  data JSONB NOT NULL,
+  UNIQUE(project, query_id, run_id, turn)
+);
+
+CREATE INDEX idx_events_project_query ON events(project, query_id);
+CREATE INDEX idx_events_run ON events(run_id);
+CREATE INDEX idx_events_type ON events(event_type);
+
+-- L0: Evidence chunks (raw, with source reference)
+CREATE TABLE IF NOT EXISTS chunks_l0 (
+  id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+  project VARCHAR NOT NULL,
+  query_id VARCHAR NOT NULL,
+  source VARCHAR NOT NULL,  -- "pi", "claude", "transcript"
+  content TEXT NOT NULL,
+  tokens INT NOT NULL,
+  created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP
+);
+
+CREATE INDEX idx_chunks_l0_project_query ON chunks_l0(project, query_id);
+
+-- L1: Per-query memories (one per standing query, up to 1024 tokens)
+CREATE TABLE IF NOT EXISTS memories_l1 (
+  id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+  project VARCHAR NOT NULL,
+  query_id VARCHAR NOT NULL,
+  content TEXT NOT NULL,
+  tokens INT NOT NULL,
+  embedding vector(768),  -- nomic-embed-text-v2-moe
+  chunks_seen INT NOT NULL,
+  chunks_used INT NOT NULL,
+  updated_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
+  run_id VARCHAR NOT NULL,
+  UNIQUE(project, query_id)
+);
+
+CREATE INDEX idx_memories_l1_project ON memories_l1(project);
+CREATE INDEX idx_memories_l1_embedding ON memories_l1 USING ivfflat (embedding vector_cosine_ops);
+
+-- L1 -> L0 provenance (which evidence chunks produced this memory)
+CREATE TABLE IF NOT EXISTS l1_l0_edges (
+  l1_id UUID REFERENCES memories_l1(id) ON DELETE CASCADE,
+  l0_id UUID REFERENCES chunks_l0(id) ON DELETE CASCADE,
+  PRIMARY KEY (l1_id, l0_id)
+);
+
+-- L2: Project synthesis (one per project, up to 1024 tokens)
+CREATE TABLE IF NOT EXISTS memories_l2 (
+  id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+  project VARCHAR NOT NULL UNIQUE,
+  content TEXT NOT NULL,
+  tokens INT NOT NULL,
+  embedding vector(768),
+  l1_count INT NOT NULL,  -- how many L1 memories were used
+  updated_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
+  run_id VARCHAR NOT NULL
+);
+
+CREATE INDEX idx_memories_l2_project ON memories_l2(project);
+CREATE INDEX idx_memories_l2_embedding ON memories_l2 USING ivfflat (embedding vector_cosine_ops);
+
+-- L2 -> L1 provenance (which L1 memories produced this synthesis)
+CREATE TABLE IF NOT EXISTS l2_l1_edges (
+  l2_id UUID REFERENCES memories_l2(id) ON DELETE CASCADE,
+  l1_id UUID REFERENCES memories_l1(id) ON DELETE CASCADE,
+  PRIMARY KEY (l2_id, l1_id)
+);
+
+-- Reference corpus (not gated, used in queries)
+CREATE TABLE IF NOT EXISTS reference_corpus (
+  id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+  project VARCHAR NOT NULL,
+  name VARCHAR NOT NULL,  -- doc name or skill name
+  content TEXT NOT NULL,
+  embedding vector(768),
+  created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
+  UNIQUE(project, name)
+);
+
+CREATE INDEX idx_corpus_project ON reference_corpus(project);
+CREATE INDEX idx_corpus_embedding ON reference_corpus USING ivfflat (embedding vector_cosine_ops);
+
+-- Ingest jobs (async queue)
+CREATE TABLE IF NOT EXISTS ingest_jobs (
+  id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+  project VARCHAR NOT NULL,
+  ingest_id VARCHAR NOT NULL UNIQUE,
+  status VARCHAR NOT NULL DEFAULT 'pending',  -- pending, processing, done, failed
+  created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
+  started_at TIMESTAMP,
+  completed_at TIMESTAMP,
+  error TEXT
+);
+
+CREATE INDEX idx_ingest_jobs_project ON ingest_jobs(project);
+CREATE INDEX idx_ingest_jobs_status ON ingest_jobs(status);
+
+-- Skills extracted from memories
+CREATE TABLE IF NOT EXISTS skills (
+  id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+  project VARCHAR NOT NULL,
+  name VARCHAR NOT NULL,
+  description TEXT NOT NULL,
+  when_to_use TEXT,
+  examples TEXT,
+  l1_source UUID NOT NULL REFERENCES memories_l1(id) ON DELETE CASCADE,
+  created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
+  UNIQUE(project, name)
+);
+
+CREATE INDEX idx_skills_project ON skills(project);

tests/it_pgvector.rs

@@ -1,76 +1,14 @@
-use mem_store::{VectorStore, VectorRecord};
+// Vector store tests now require PostgreSQL connection
+// See tests with database fixtures or use integration tests
 
 #[test]
-fn a1_insert_and_search() {
+#[ignore]
-    let mut store = VectorStore::new();
+fn _vector_search_requires_database() {
-    
+    // VectorStore is now backed by PostgreSQL with pgvector extension
-    // Insert two similar vectors
+    // Tests require:
-    let v1 = vec![1.0, 0.0, 0.0];
+    // - Running CNPG cluster
-    let v2 = vec![0.99, 0.1, 0.0];
+    // - Database initialized with schema
-    let v3 = vec![0.0, 0.0, 1.0];  // orthogonal
+    // - Connection pooling setup
-    
+    //
-    store.insert(VectorRecord {
+    // Use integration tests with database containers for full testing
-        id: "r1".to_string(),
-        chunk_id: "c1".to_string(),
-        kind: "text".to_string(),
-        embedding: v1,
-        tokens: 100,
-    }).unwrap();
-    
-    store.insert(VectorRecord {
-        id: "r2".to_string(),
-        chunk_id: "c2".to_string(),
-        kind: "text".to_string(),
-        embedding: v2,
-        tokens: 100,
-    }).unwrap();
-    
-    store.insert(VectorRecord {
-        id: "r3".to_string(),
-        chunk_id: "c3".to_string(),
-        kind: "text".to_string(),
-        embedding: v3,
-        tokens: 100,
-    }).unwrap();
-    
-    // Search for vectors similar to v1
-    let results = store.search(&[1.0, 0.0, 0.0], 3, 0.0).unwrap();
-    
-    // r1 should be first (identical)
-    assert_eq!(results[0].0, "r1");
-    assert!((results[0].1 - 1.0).abs() < 0.01);
-    
-    // r2 should be second (similar)
-    assert_eq!(results[1].0, "r2");
-    assert!(results[1].1 > 0.9);
-    
-    // r3 should be last (orthogonal)
-    assert_eq!(results[2].0, "r3");
-    assert!(results[2].1 < 0.1);
-}
-
-#[test]
-fn a2_min_score_filter() {
-    let mut store = VectorStore::new();
-    
-    store.insert(VectorRecord {
-        id: "r1".to_string(),
-        chunk_id: "c1".to_string(),
-        kind: "text".to_string(),
-        embedding: vec![1.0, 0.0],
-        tokens: 100,
-    }).unwrap();
-    
-    store.insert(VectorRecord {
-        id: "r2".to_string(),
-        chunk_id: "c2".to_string(),
-        kind: "text".to_string(),
-        embedding: vec![0.0, 1.0],
-        tokens: 100,
-    }).unwrap();
-    
-    // Search with high threshold - only perfect match
-    let results = store.search(&[1.0, 0.0], 10, 0.99).unwrap();
-    assert_eq!(results.len(), 1);
-    assert_eq!(results[0].0, "r1");
 }