feat(s032): complete slice 032 (semantic memory and hybrid retrieval)

.github/workflows/gate.yml

@@ -60,7 +60,9 @@ jobs:
     runs-on: ubuntu-latest
     services:
       postgres:
-        image: postgres:17
+        # Slice 032: pgvector's image of Postgres 17, so the vector extension exists for the
+        # semantic memory's Postgres store.
+        image: pgvector/pgvector:pg17
         env:
           POSTGRES_USER: trinity
           POSTGRES_PASSWORD: trinity

.github/workflows/package.yml

@@ -192,12 +192,20 @@ jobs:
           ours() { ps -eo pid,ppid,comm | grep -E 'burrito|beam\.smp|erl_child_setup|desktop_' | grep -v grep || true; }
           ours > ps-before.txt
           chmod +x "burrito_out/${{ matrix.artifact }}"
-          "./burrito_out/${{ matrix.artifact }}" --no-halt --smoke | tee smoke.log
+          # Slice 032: the variable, not `--smoke`: Kernel.CLI read the argument as a file
+          # ("No file named --smoke", exit 1) and won the race against the halting Task on
+          # macOS in run 35608084951.
+          TRINITY_SMOKE=1 "./burrito_out/${{ matrix.artifact }}" --no-halt | tee smoke.log
           sleep 3
           ours > ps-after.txt
           grep -q '^TRINITY_SMOKE_PORT=[0-9][0-9]*$' smoke.log
           # Slice 013: the renderer's NIF loaded and rendered in this binary (exit 3 otherwise).
           grep -q '^TRINITY_SMOKE_MARKDOWN=ok$' smoke.log
+          # Slice 032, AC7: a fake-vector search ran inside this binary through the vector
+          # store in force (exit 4 otherwise). The EXLA and SEMANTIC lines are recorded, not
+          # asserted: whether the XLA library loads in Burrito's ERTS is what they measure.
+          grep -q '^TRINITY_SMOKE_VEC=ok:' smoke.log
+          grep -E '^TRINITY_SMOKE_(EXLA|SEMANTIC)=' smoke.log
           echo "--- ours, before ---"; cat ps-before.txt
           echo "--- ours, after ----"; cat ps-after.txt
           diff ps-before.txt ps-after.txt
@@ -207,14 +215,20 @@ jobs:
         shell: pwsh
         run: |
           $ErrorActionPreference = "Stop"
-          & ".\burrito_out\${{ matrix.artifact }}" --no-halt --smoke | Tee-Object smoke.log
+          $env:TRINITY_SMOKE = "1"
+          & ".\burrito_out\${{ matrix.artifact }}" --no-halt | Tee-Object smoke.log
           if ($LASTEXITCODE -ne 0) { throw "smoke exited $LASTEXITCODE" }
           if (-not (Select-String -Path smoke.log -Pattern '^TRINITY_SMOKE_PORT=\d+$')) {
             throw "no port line"
           }
           if (-not (Select-String -Path smoke.log -Pattern '^TRINITY_SMOKE_MARKDOWN=ok$')) {
             throw "the markdown renderer did not render in this binary (slice 013)"
           }
+          # Slice 032, AC7, as on the other two: the vector search binds, the other two lines are recorded.
+          if (-not (Select-String -Path smoke.log -Pattern '^TRINITY_SMOKE_VEC=ok:')) {
+            throw "the vector search did not run inside this binary (slice 032)"
+          }
+          Select-String -Path smoke.log -Pattern '^TRINITY_SMOKE_(EXLA|SEMANTIC)='
 
       # AC1's property on a runner, and AC6's per-OS cold start, which used to be skipped.
       # `bash` on every OS, Windows included: the job that never curled was the one gap in

ROADMAP.md

@@ -46,7 +46,7 @@ standards register names the rows that ask for them.
 | 026 | Store-and-forward receipts for disconnected operation | 2 Tools | L | 024 | planned |
 | 030 | Persona (SOUL) + always-on memory tier | 3 Memory | M | 012 | approved |
 | 031 | Session search (SQLite FTS5) | 3 Memory | S | 010 | approved |
-| 032 | Embeddings + semantic memory + hybrid retrieval | 3 Memory | L | 031 | planned |
+| 032 | Embeddings + semantic memory + hybrid retrieval | 3 Memory | L | 031 | done |
 | 033 | Project context: AGENTS.md | 3 Memory | S | 030, 022 | approved |
 | 034 | Export, import, restore | 3 Memory | S | 030, 031 | approved |
 | 040 | Skills registry + agentskills.io format + progressive disclosure | 4 Skills | M | 020 | planned |

VERSIONS.md

@@ -87,7 +87,7 @@ never pin a version hex marks as retired or vulnerable.
 | `ecto_sql` | ~> 3.13 | ✅ in `mix.lock` |  |
 | `ecto_sqlite3` | >= 0.0.0 | ✅ in `mix.lock` | Primary DB. FTS5 available. |
 | `postgrex` | >= 0.0.0 (optional) | ✅ in `mix.lock` | Secondary DB driver, `optional: true` so the desktop build carries none of it; compiled in only under `TRINITY_DB=postgres`, which the CI job proves. Added at Slice 010. Was one row with pgvector; pgvector keeps its own row below. |
-| `pgvector` | optional, ~> 0.3 | 🔍 not yet a dependency | Vectors on the Postgres path. Not yet a dependency; Slice 032 decides. Split from the postgrex row at Slice 010. |
+| `pgvector` | ~> 0.4.1 | ✅ in `mix.lock` | Vectors on the Postgres path: `memories.embedding_vector vector(384)` under an HNSW cosine index, searched by `Trinity.Memory.VectorStores.Pgvector` (Slice 032). The postgres job runs on the `pgvector/pgvector:pg17` image. Split from the postgrex row at Slice 010. |
 | `oban` | ~> 2.24 | 🔍 not yet a dependency | Uses `Oban.Engines.Lite` on SQLite. ⚠️ Oban Pro Workflows/Smart engine are Postgres-only. Added at Slice 050. |
 | `req` | ~> 0.5 | ✅ in `mix.lock` | HTTP client. |
 | `req_llm` | ~> 1.22 | ✅ in `mix.lock` | Provider layer (streaming, tools, structured output, usage). ⚠️ The pin was `~> 1.10` against a recorded latest of 1.10.0; the real latest was twelve minors ahead. Check event shapes against the current version at Slice 011, not against this file's prose. Added at Slice 011. |
@@ -118,10 +118,11 @@ never pin a version hex marks as retired or vulnerable.
 
 | Name | Pin | Verified | Note |
 |---|---|---|---|
-| `nx, exla` | latest stable | 🔍 not a single package | Local embeddings. EXLA binary size matters for desktop: measure in 032. Two packages, so no single lock key. |
-| `bumblebee` | ~> 0.7 | 🔍 not yet a dependency | `all-MiniLM-L6-v2` embeddings; Whisper later. Added at Slice 032. |
-| `sqlite_vec` | ~> 0.1 | 🔍 not yet a dependency | Vectors in SQLite. Verify the loadable extension works inside the Burrito bundle (Slice 032). ⚠️ Pre-1.0, no release in roughly 22 months, 6,938 downloads all-time. R11's trigger already fires. Decide the fallback before Slice 032 starts. |
-| `hnswlib` | ~> 0.1.7 | 🔍 not yet a dependency | ⚠️ Pre-1.0. Optional accelerator; not on the critical path. |
+| `nx` | ~> 0.13.1 | ✅ in `mix.lock` | Tensors for the local embedder (Slice 032). The 0.13 line, not 1.0.0 (2026-09-10): bumblebee 0.7.1 accepts `~> 0.12 or ~> 0.13` and has no release for 1.0. Measured at Slice 032: docs/perf.md. |
+| `exla` | ~> 0.13.1 | ✅ in `mix.lock` | The XLA backend the serving compiles to (Slice 032); precompiled for x86_64 and aarch64 Linux and macOS, none for Windows (not declared on a Windows host; the tier is off there). Declared `runtime: false` and carried by the release in `:load` mode, started on demand by `Trinity.Memory.Embedders.Bumblebee.exla/0`: its NIF does not load in Burrito's Linux ERTS (NOTES finding 2). The XLA shared library is 463 MB on disk: docs/perf.md. Pulls `xla` 0.10.0 and `fine`. |
+| `bumblebee` | ~> 0.7.1 | ✅ in `mix.lock` | `all-MiniLM-L6-v2` embeddings through `Trinity.Memory.Embedders.Bumblebee` (Slice 032); Whisper later. Pulls `axon` 0.8.1 and `tokenizers` 0.5.1 (a precompiled Rust NIF). |
+| `sqlite_vec` | not used (decided at Slice 032, 2026-09-21) | 🔍 not a single package | Vectors in SQLite. Its only release (0.1.0, 2024-11-19) requires `nx ~> 0.9` and cannot sit in a tree with nx 0.13; R11's trigger fired and the owner chose brute force in Elixir over the persona's semantic rows (`Trinity.Memory.VectorStores.Brute`, docs/02: fine to about 10^5) with pgvector on the Postgres path. The row stays so the decision is visible where a reader would look for the package. |
+| `hnswlib` | ~> 0.1.10 | 🔍 not yet a dependency | ⚠️ Pre-1.0 (0.1.10, 2026-09-17, active). The scale option after brute force; not yet a dependency, the slice that measures brute force past 10^5 rows adds it. |
 
 ### Tools, sandbox and desktop
 

config/config.exs

@@ -73,6 +73,8 @@ config :trinity, :tools,
     Trinity.Tools.Web.Search,
     # Slice 031: full-text search over past messages.
     Trinity.Tools.SessionSearch,
+    # Slice 032: hybrid recall over semantic memories and past messages.
+    Trinity.Tools.Recall,
     # Slice 030: the always-on memory tiers.
     Trinity.Tools.Memory,
     Trinity.Tools.Shell.Run
@@ -82,7 +84,7 @@ config :trinity, :tools,
     web: ["web_fetch", "web_search"],
     shell: ["shell"],
     # Slice 031: search over past conversations.
-    memory: ["session_search", "memory"]
+    memory: ["session_search", "recall", "memory"]
   }
 
 # Slice 022: the filesystem roots beside the data directory (always a root) and the session's

config/runtime.exs

@@ -24,7 +24,9 @@ import Config
 # which is the same call: this file is evaluated before the application is available under
 # Mix, so it may not call project modules. `Burrito.Util.Args.get_arguments/0` reads the same
 # thing.
-smoke? = "--smoke" in Enum.map(:init.get_plain_arguments(), &to_string/1)
+smoke? =
+  "--smoke" in Enum.map(:init.get_plain_arguments(), &to_string/1) or
+    System.get_env("TRINITY_SMOKE") == "1"
 
 if System.get_env("PHX_SERVER") || smoke? do
   config :trinity, TrinityWeb.Endpoint, server: true
@@ -119,7 +121,9 @@ if config_env() == :prod do
   if System.get_env("TRINITY_DB") == "postgres" do
     config :trinity, Trinity.Repo,
       url: System.get_env("DATABASE_URL") || raise("TRINITY_DB=postgres needs DATABASE_URL"),
-      pool_size: String.to_integer(System.get_env("POOL_SIZE") || "10")
+      pool_size: String.to_integer(System.get_env("POOL_SIZE") || "10"),
+      # Slice 032: pgvector's `vector` type.
+      types: Trinity.Repo.PostgrexTypes
 
     config :trinity, Trinity.Repo.Receipts,
       url: System.get_env("DATABASE_URL"),

config/test.exs

@@ -27,6 +27,8 @@ config :trinity, :tools,
     Trinity.Tools.Web.Search,
     # Slice 031: full-text search over past messages.
     Trinity.Tools.SessionSearch,
+    # Slice 032: hybrid recall.
+    Trinity.Tools.Recall,
     # Slice 030: the always-on memory tiers.
     Trinity.Tools.Memory,
     Trinity.Tools.Shell.Run
@@ -37,7 +39,7 @@ config :trinity, :tools,
     web: ["web_fetch", "web_search"],
     shell: ["shell"],
     # Slice 031: search over past conversations.
-    memory: ["session_search", "memory"]
+    memory: ["session_search", "recall", "memory"]
   },
   timeout_ms: 2_000
 
@@ -97,21 +99,36 @@ if System.get_env("TRINITY_DB") == "postgres" do
   config :trinity, Trinity.Repo,
     url: System.get_env("DATABASE_URL") || raise("TRINITY_DB=postgres needs DATABASE_URL"),
     pool_size: 10,
-    pool: Ecto.Adapters.SQL.Sandbox
+    pool: Ecto.Adapters.SQL.Sandbox,
+    # Slice 032: pgvector's `vector` type.
+    types: Trinity.Repo.PostgrexTypes,
+    # Slice 032: the sandbox shares one connection with every process a test starts, and
+    # DBConnection sheds load (refuses checkouts) when the queue stays over `queue_target`
+    # for `queue_interval`. The 100-session and 20-writer tests queue that hard on the
+    # postgres job's runner, more so once each turn also runs the retriever (runs 35610164189
+    # and 35610171809); a queue that is long on purpose should wait, not shed.
+    queue_target: 5_000,
+    queue_interval: 30_000
 
   # Slice 024: the receipts Repo shares the Postgres database (its own migrations table).
   config :trinity, Trinity.Repo.Receipts,
     url: System.get_env("DATABASE_URL"),
     pool_size: 10,
-    pool: Ecto.Adapters.SQL.Sandbox
+    pool: Ecto.Adapters.SQL.Sandbox,
+    queue_target: 5_000,
+    queue_interval: 30_000
 else
   config :trinity, Trinity.Repo,
     database: Path.expand("../trinity_test.db", __DIR__),
-    pool: Ecto.Adapters.SQL.Sandbox
+    pool: Ecto.Adapters.SQL.Sandbox,
+    queue_target: 5_000,
+    queue_interval: 30_000
 
   config :trinity, Trinity.Repo.Receipts,
     database: Path.expand("../trinity_test_receipts.db", __DIR__),
-    pool: Ecto.Adapters.SQL.Sandbox
+    pool: Ecto.Adapters.SQL.Sandbox,
+    queue_target: 5_000,
+    queue_interval: 30_000
 end
 
 # We don't run a server during test. If one is required,
@@ -150,3 +167,7 @@ config :phoenix,
 # Slice 024: the receipt signing key and registry for the suite live under the project's
 # ignored tmp/, never in the data directory of the machine running the tests.
 config :trinity, :receipts, keys_dir: Path.expand("../tmp/test_keys", __DIR__)
+
+# Slice 032: the suite embeds with the deterministic fake; nothing leaves the machine and no
+# model is needed.
+config :trinity, :memory, embedder: :fake, observer: false

coverage.tsv

@@ -15,3 +15,4 @@ slice_id	percent	sha	date
 030	78.40	c657da8	2026-09-21
 033	78.50	3b5e89e	2026-09-21
 034	78.83	3a5785f	2026-09-21
+032	78.99	193800e	2026-09-21

docs/01-architecture.md

@@ -42,7 +42,12 @@ Trinity.Application
 │                 # INSERT but does not guarantee each row read the correct predecessor.
 ├── Trinity.Effects.Boot                          # Slice 024, as built: a transient Task writing the boot receipt
 │                                                 # once the signer and the authority are known
-├── Trinity.Memory.Supervisor                     # Nx.Serving for embeddings, retrieval. Slice 032
+├── Trinity.Memory.Supervisor                     # Slice 032, as built: Trinity.Memory.TaskSupervisor (the
+│   │                                             # observer's task after each turn, the page's model download)
+│   └── Nx.Serving (Trinity.Memory.Embedding)     # the local embedder, started only when the embedder is
+│                                                 # :local and the model is in the cache; absent otherwise, the
+│                                                 # tier off and the reason logged once. ensure_embedding/0
+│                                                 # starts it after a download.
 ├── Trinity.Skills.Registry                       # hot-loaded skills index. Slice 040
 ├── Oban                                        # cron + durable jobs. Slice 050
 ├── Trinity.MCP.Supervisor                        # MCP clients (one per server, lib per 059) + server. Slice 059-062
@@ -106,7 +111,8 @@ through those modules only. `boundary` `exports:` lists enforce this.
 | `Trinity.LLM.Provider` | `stream/3`, `generate/3`, `embed/2`, `models/0`, `capabilities/1` | config `:providers` list |
 | `Trinity.Tools.Tool` | `name/0`, `description/0`, `schema/0`, `risk/0`, `execute/2` | config `:tools` list + MCP dynamic registration |
 | `Trinity.Gateways.Adapter` | `child_spec/1`, `deliver/2`, `capabilities/0` | config `:gateways` list |
-| `Trinity.Memory.VectorStore` | `upsert/2`, `search/3`, `delete/1` | config `:vector_store` (`SqliteVec` \| `Pgvector` \| `Hnswlib`) |
+| `Trinity.Memory.VectorStore` | `upsert/3`, `search/3`, `delete/1`, `count/1` (as built at 032: the scope filter is `search/3`'s required argument) | the database adapter, at compile time (`Brute` on SQLite, `Pgvector` on Postgres; `Hnswlib` when a slice measures past 10^5 rows) |
+| `Trinity.Memory.Embedder` | `embed/1`, `dim/0`, `model_id/0`, `availability/0` | `config :trinity, :memory, embedder:` (`:local` default, `:hosted` opt-in only, `:fake` in the suite) |
 | `Trinity.Skills.Loader` | `load/1`, `validate/1` | config `:skill_loaders` |
 | `Trinity.Permissions.Policy` | `decide/3` → `:allow | :deny | {:ask, prompt}` | config |
 
@@ -120,7 +126,7 @@ UI/Gateway ──user_message──▶ Session(gen_statem)
   Session: append message → build prompt (persona + always-on memory + skills index + history[compacted])
           → LLM.stream/3 ──chunks──▶ PubSub "session:<id>" ──▶ LiveView + gateways
           → on tool_call: Permissions.decide → (ask → approval_wait) → Tools.execute in Task → result appended
-          → loop until final text → persist → Memory.observe(turn) (async) → idle
+          → loop until final text → persist → Memory.Observer.observe(turn) (a Task under Trinity.Memory.TaskSupervisor, as built at 032) → idle
 ```
 
 Every state transition is persisted before it is broadcast. A crash between persist and broadcast is safe

docs/02-tech-stack.md

@@ -14,6 +14,7 @@ Versions live in `VERSIONS.md`. This file explains *why* each choice was made an
 | MCP | decided by Slice 059 (fastest_mcp / gen_mcp / anubis_mcp / own server) | Target is 2026-07-28 with 2025-11-25 compat; anubis is ≤ 2025-11-25 and LGPL-3.0 | ADR-0007 finalised by measurement |
 | Embeddings | Bumblebee + EXLA (all-MiniLM-L6-v2) | Local, private, 384-dim | If EXLA binary size is unacceptable on desktop → hosted embeddings via req_llm, or Ortex (risk: stalled) |
 | Vector search | sqlite_vec (brute force) behind `VectorStore` behaviour | Fine to ~10^5 vectors; no extra process | Scale → hnswlib (pre-1.0) or pgvector HNSW |
+| Vector search, as built at 032 | brute force in Elixir over the row's `embedding` bytes (SQLite); pgvector HNSW (Postgres) | sqlite_vec's one release needs `nx ~> 0.9` and cannot share a tree with nx 0.13 (VERSIONS.md); measured in docs/perf.md | Scale → hnswlib 0.1.10 |
 | Shell tool | MuonTrap | Guaranteed child kill on process death; cgroups on Linux | none |
 | Sandbox | Luerl (`sandbox` pkg) | In-VM, reduction-limited, no OS access | Untrusted native code → container/microVM (out of scope) |
 | Modularity | `boundary` + behaviours + `Registry` | Compile-time enforcement of context deps | none |

docs/05-data-model.md

@@ -75,6 +75,22 @@ vocabulary is `global | persona:<id> | project:<path> | session:<id>` (a session
 `persona`, `global`), `key` is required for the two always-on tiers (`^[a-z0-9][a-z0-9_.-]{0,63}$`, unique with
 tier and scope), and the semantic tier's columns wait for 032. Two tables beside it:
 
+As built at slice 032 (the semantic tier): no `memories_vec` virtual table and no `sqlite_vec` (its only release
+needs `nx ~> 0.9`; NOTES decision 1). The vector is three columns on the row, on both databases:
+`embedding` (binary: float32, little-endian, `dim * 4` bytes), `embedding_model` (the embedder that produced it,
+e.g. `bumblebee:sentence-transformers/all-MiniLM-L6-v2`, `fake:sha256-384` in the suite) and `embedding_dim`
+(integer), all nullable, under an index on (`persona_id`, `tier`, `embedding_model`). On Postgres the migration
+also creates the `vector` extension, a fourth column `embedding_vector vector(384)` and an HNSW cosine index on
+it where `tier = 'semantic'`; `Trinity.Memory.VectorStores.Pgvector` keeps it in step with the bytes column and
+searches it with `<=>`, and `VectorStores.Brute` (SQLite) loads the filter's rows and scores them in Elixir.
+A search takes the scope list as a required argument (M6) and runs only over rows whose `embedding_model` is the
+embedder in force's: models are never mixed, a change of embedder is a re-embed. Semantic rows are written by
+`Trinity.Memory.Semantic.add/2` (the observer after a turn, `by: "observer"`; the page, `by: "ui"`), keyed
+`<slug of the body>-<6 hex of its SHA-256>`, and every write is a `memory_changes` row (`add`, `remove`, `pin`).
+The change log's `action` vocabulary grows by `pin` (a semantic memory promoted to `always_on` through
+`AlwaysOn.add/2`, so the budget applies) and `by` by `observer`. Recall marks its memory hits' `last_used_at`,
+which the retriever's recency decay reads.
+
 ### memory_changes (Slice 030)
 `persona_id`, `action` (add | replace | remove | promote | consolidate), `tier`, `scope`, `key`, `before`,
 `after`, `by` (tool | ui | consolidator), `session_id`, `proposal_id`, `inserted_at`. Every write to the