docs: update specs and changelog for v2026.3.70 (T134 Brain Memory Automation)

- CLEO-API.md: Section 15.10 — new CLI commands, config section, contract
  types, internal exports, @xenova/transformers dependency
- CORE-PACKAGE-SPEC.md: Updated memory module description, config types,
  runtime dependencies, new Section 15.9 for T134
- CLEOOS-VISION.md: Updated Brain Intelligence section with shipped features
  (vector similarity, summarization, bridge automation, maintenance, transcript)
- CHANGELOG.md: v2026.3.70 entry with all 12 T134 tasks

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: kryptobaseddev

CHANGELOG.md

@@ -2,6 +2,26 @@
 
 ## [Unreleased]
 
+## [2026.3.70] — 2026-03-23
+
+### Added
+- **Brain Memory Automation** (T134 epic, 12 tasks):
+  - `BrainConfig` typed configuration section with defaults and templates (T135)
+  - Local embedding provider via `@xenova/transformers` all-MiniLM-L6-v2, dynamic import (T136)
+  - Embedding worker thread + async queue for non-blocking processing (T137)
+  - Memory bridge refresh wired to lifecycle hooks with 30s debounce (T138)
+  - Context-aware memory bridge generation using `hybridSearch()` + token budget (T139)
+  - Session summarization: dual-mode prompt + structured `SessionSummaryInput` response (T140)
+  - Auto-link observations to focused task via `brain_memory_links` (T141)
+  - Embedding backfill with progress reporting: `cleo backfill --embeddings` (T142)
+  - Brain maintenance command: `cleo brain maintenance` with `--skip-decay`, `--skip-consolidation`, `--skip-embeddings` (T143)
+  - Cross-provider transcript hook on `AdapterHookProvider` + Claude Code adapter implementation (T144)
+  - Updated CLEO-INJECTION.md templates with Memory Automation section (T145)
+  - Updated CLEO-BRAIN-SPECIFICATION.md to v2.0.0 (T146)
+
+### Dependencies
+- Added `@xenova/transformers` ^2.17.2 to `@cleocode/core` (external, dynamic import)
+
 ## [2026.3.69] - 2026-03-23
 
 ### Fixed

docs/concepts/CLEOOS-VISION.md

@@ -147,7 +147,11 @@ The BRAIN system in `@cleocode/core` stores observations, patterns, learnings, a
 
 - **Three-layer retrieval** (shipped): search/timeline/fetch with ~10x token savings over traditional RAG
 - **FTS5 full-text search** (shipped): keyword search across all brain tables
-- **Vector similarity** (in progress): SQLite-vec integration for semantic retrieval
+- **Vector similarity** (shipped): Local embedding via `@xenova/transformers` all-MiniLM-L6-v2 — dynamic import, zero overhead when disabled (T136)
+- **Session summarization** (shipped): Dual-mode prompt + structured `SessionSummaryInput` ingestion on `session.end` (T140)
+- **Memory bridge automation** (shipped): Hook-driven refresh on `session.end` and `tasks.complete` with 30-second debounce; context-aware generation via `hybridSearch()` within configurable token budget (T138, T139)
+- **Brain maintenance** (shipped): `cleo brain maintenance` combines temporal decay + consolidation + embedding backfill into a single governed command (T143)
+- **Cross-provider transcript extraction** (shipped): Optional `getTranscript()` on `AdapterHookProvider`; Claude Code adapter implementation auto-extracts observation candidates from session transcripts (T144)
 - **Knowledge graph** (planned): relationship-based discovery (updates/extends/derives) with temporal reasoning
 - **Active circulation** (planned): Living BRAIN reinforces useful memories, detects contradictions, and surfaces relevant context proactively
 
@@ -240,6 +244,7 @@ The kernel is shipped and operational:
 - Task work on facade: `tasks.start/stop/current` — no direct barrel imports needed (T126)
 - Bootstrap injection chain: legacy template sync, CAAMP sanitization, post-bootstrap health check (T124)
 - Migration resilience: journal reconciliation and `ensureRequiredColumns()` safety net (v2026.3.61)
+- Brain Memory Automation (T134 epic, v2026.3.70): `BrainConfig` typed section, local embedding via `@xenova/transformers` all-MiniLM-L6-v2, async embedding queue, lifecycle-driven bridge refresh with debounce, context-aware bridge generation, session summarization (dual-mode), auto-link observations to focused task, embedding backfill CLI (`cleo backfill --embeddings`), brain maintenance command (`cleo brain maintenance`), cross-provider transcript extraction via adapter hook, updated injection templates and CLEO-BRAIN-SPECIFICATION.md v2.0.0
 
 ### What Is Specified (Runtime, Conduit)
 

docs/specs/CLEO-API.md

@@ -557,6 +557,69 @@ Sixteen GitHub issues (#63–#78) resolved across five point releases. Key fixes
 - `paginate()` null guard (v2026.3.64)
 - `detect-drift` user project detection (v2026.3.65)
 
+### 15.10 T134 — Brain Memory Automation (v2026.3.70)
+
+The T134 epic (12 tasks, T135–T146) delivers the first full automation layer for BRAIN memory: local embeddings, lifecycle-driven bridge refresh, session summarization, cross-provider transcript extraction, and a combined maintenance command.
+
+#### New CLI Commands (T134)
+
+| Command | Purpose |
+|---------|---------|
+| `cleo brain maintenance` | Combined brain maintenance: temporal decay + consolidation + embedding backfill (T143) |
+| `cleo brain maintenance --skip-decay` | Skip temporal decay step |
+| `cleo brain maintenance --skip-consolidation` | Skip memory consolidation step |
+| `cleo brain maintenance --skip-embeddings` | Skip embedding backfill step |
+| `cleo backfill --embeddings` | Retroactive embedding backfill with progress reporting (T142) |
+
+#### New Config Section (T135)
+
+The `brain` section is now a first-class typed config block in `CleoConfig`:
+
+| Key | Type | Default | Purpose |
+|-----|------|---------|---------|
+| `brain.autoCapture` | `boolean` | `true` | Auto-capture observations on lifecycle events |
+| `brain.embedding.enabled` | `boolean` | `false` | Enable local embedding provider |
+| `brain.embedding.provider` | `'local' \| 'openai' \| 'custom'` | `'local'` | Embedding provider selection |
+| `brain.memoryBridge.autoRefresh` | `boolean` | `true` | Auto-refresh bridge on session/task lifecycle events |
+| `brain.memoryBridge.contextAware` | `boolean` | `false` | Use `hybridSearch()` for context-aware bridge content |
+| `brain.memoryBridge.maxTokens` | `number` | `2000` | Token budget for bridge content |
+| `brain.summarization.enabled` | `boolean` | `false` | Enable session summarization on `session.end` |
+
+#### New Contract Types (T134)
+
+| Type | Purpose |
+|------|---------|
+| `BrainConfig` | Top-level brain config section in `CleoConfig` |
+| `BrainEmbeddingConfig` | `brain.embedding.*` sub-interface |
+| `BrainMemoryBridgeConfig` | `brain.memoryBridge.*` sub-interface |
+| `BrainSummarizationConfig` | `brain.summarization.*` sub-interface |
+| `SessionSummaryInput` | Structured session summary type for ingestion |
+
+**Updated interfaces:**
+
+- `AdapterHookProvider` gains an optional `getTranscript()` method for cross-provider transcript extraction (T144)
+- `SessionEndResult` gains `memoryPrompt?: string` for dual-mode summarization
+- `SessionEndParams` gains `sessionSummary?: SessionSummaryInput` for structured summary ingestion
+
+#### New Internal Exports (T134)
+
+| Export | Source | Purpose |
+|--------|--------|---------|
+| `initDefaultProvider` | `memory/brain-embedding.ts` | Initialize local all-MiniLM-L6-v2 embedding provider (T136) |
+| `generateContextAwareContent` | `memory/memory-bridge.ts` | Scope-aware bridge generation using `hybridSearch()` (T139) |
+| `buildSummarizationPrompt` | `memory/session-memory.ts` | Build agent summarization prompt for dual-mode output (T140) |
+| `ingestStructuredSummary` | `memory/session-memory.ts` | Auto-ingest structured `SessionSummaryInput` to brain.db (T140) |
+| `extractFromTranscript` | `memory/auto-extract.ts` | Extract observations from provider transcript via hook (T144) |
+| `runBrainMaintenance` | `memory/brain-maintenance.ts` | Combined maintenance: decay + consolidation + embedding backfill (T143) |
+| `LocalEmbeddingProvider` | `memory/embedding-local.ts` | all-MiniLM-L6-v2 ONNX embedding via `@xenova/transformers` (T136) |
+| `EmbeddingQueue` | `memory/embedding-queue.ts` | Async embedding queue for non-blocking background processing (T137) |
+
+#### New Dependency (T136)
+
+| Package | Version | Load Strategy |
+|---------|---------|---------------|
+| `@xenova/transformers` | `^2.17.2` | Dynamic import — only loads when `brain.embedding.enabled: true` |
+
 ---
 
 ## 9. Document Hierarchy

docs/specs/CORE-PACKAGE-SPEC.md

@@ -161,7 +161,7 @@ The public barrel (`packages/core/src/index.ts`) re-exports all public modules a
 | `issue` | `packages/core/src/issue/` | Issue and bug tracking | Yes |
 | `lib` | `packages/core/src/lib/` | General-purpose utilities: `withRetry` (exponential backoff), `computeDelay`. No database coupling. | No |
 | `lifecycle` | `packages/core/src/lifecycle/` | RCASD-IVTR+C gate enforcement, stage transitions | Yes (SQLite lifecycle tables) |
-| `memory` | `packages/core/src/memory/` | Brain.db observations, search, 3-layer retrieval | No (uses brain.db directly) |
+| `memory` | `packages/core/src/memory/` | Brain.db observations, search, 3-layer retrieval, **local embedding provider, async embedding queue, context-aware bridge, session summarization, brain maintenance** | No (uses brain.db directly) |
 | `metrics` | `packages/core/src/metrics/` | Telemetry, value tracking, provider detection | No |
 | `migration` | `packages/core/src/migration/` | Schema version detection and migration execution | Yes (SQLite) |
 | `nexus` | `packages/core/src/nexus/` | Cross-project registry operations (nexus.db) | No (uses nexus.db) |
@@ -405,7 +405,7 @@ The public barrel re-exports all types from `@cleocode/contracts` via `export *
 | LAFS types | `LafsSuccess`, `LafsError`, `LafsEnvelope`, `LAFSMeta`, `LAFSPage`, `Warning`, `MVILevel`, `GatewayEnvelope` |
 | DataAccessor | `DataAccessor`, `TaskQueryFilters`, `QueryTasksResult`, `TaskFieldUpdates`, `TransactionAccessor` |
 | Exit codes | `ExitCode`, `isErrorCode`, `isSuccessCode`, `getExitCodeName` |
-| Config types | `CleoConfig`, `HierarchyConfig`, `SessionConfig`, `LifecycleConfig`, `BackupConfig`. Note: `enforcement.*` and `verification.*` sections are live at runtime but not yet declared in `CleoConfig` — they are read via untyped dot-path. The T101 config audit reduced the declared schema from ~283 to ~113 fields. |
+| Config types | `CleoConfig`, `HierarchyConfig`, `SessionConfig`, `LifecycleConfig`, `BackupConfig`, `BrainConfig`, `BrainEmbeddingConfig`, `BrainMemoryBridgeConfig`, `BrainSummarizationConfig`. Note: `enforcement.*` and `verification.*` sections are live at runtime but not yet declared in `CleoConfig` — they are read via untyped dot-path. The T101 config audit reduced the declared schema from ~283 to ~113 fields. The T135 task added the `brain` section as a fully typed block in `CleoConfig` with four sub-interfaces. |
 | Status registry | `TASK_STATUSES`, `SESSION_STATUSES`, `STATUS_REGISTRY`, `isValidStatus` |
 | Adapter types | `CLEOProviderAdapter`, `AdapterManifest`, `DetectionPattern`, `AdapterCapabilities` |
 | WarpChain types | `WarpChain`, `WarpChainInstance`, `WarpStage`, `ChainShape`, `GateContract` |
@@ -566,6 +566,7 @@ The following are internal implementation details, not part of the public contra
 | `write-file-atomic` | `^6.0.0` | Atomic file write operations |
 | `yaml` | `^2.8.2` | YAML parsing and serialization |
 | `zod` | `^3.25.76` | Runtime validation schemas (used via drizzle-orm Zod integration) |
+| `@xenova/transformers` | `^2.17.2` | Local ONNX embedding (all-MiniLM-L6-v2), dynamic import — only loads when `brain.embedding.enabled: true` | No |
 
 ### 9.2 Dev Dependencies
 
@@ -874,6 +875,76 @@ Sixteen GitHub issues (#63–#78) resolved across five point releases:
 | `paginate()` null guard | v2026.3.64 | Handles undefined/null input arrays |
 | `detect-drift` user projects | v2026.3.65 | Distinguishes CLEO source repo from user projects |
 
+### 15.9 New Features in T134 Brain Memory Automation (v2026.3.70)
+
+The T134 epic (12 tasks, T135–T146) delivers the BRAIN memory automation layer: typed config, local embeddings, lifecycle-driven bridge refresh, session summarization, cross-provider transcript extraction, and a combined maintenance command.
+
+#### T135 — BrainConfig Typed Configuration
+
+The `brain` section is now a first-class typed block in `CleoConfig` with four sub-interfaces declared in `@cleocode/contracts`:
+
+| Interface | Purpose |
+|-----------|---------|
+| `BrainConfig` | Top-level `brain` config section |
+| `BrainEmbeddingConfig` | `brain.embedding.*` settings |
+| `BrainMemoryBridgeConfig` | `brain.memoryBridge.*` settings |
+| `BrainSummarizationConfig` | `brain.summarization.*` settings |
+
+Config templates ship with strict defaults (`autoCapture: true`, `embedding.enabled: false`, `memoryBridge.autoRefresh: true`).
+
+#### T136 — Local Embedding Provider
+
+`LocalEmbeddingProvider` in `memory/embedding-local.ts` provides all-MiniLM-L6-v2 embeddings via `@xenova/transformers`. The dependency is dynamically imported — zero overhead unless `brain.embedding.enabled: true`. Exposed via internal barrel as `initDefaultProvider`.
+
+#### T137 — Async Embedding Queue
+
+`EmbeddingQueue` in `memory/embedding-queue.ts` provides a non-blocking worker-thread queue for background embedding generation. Observations written to brain.db are queued for embedding without blocking the write path.
+
+#### T138 — Memory Bridge Lifecycle Hooks
+
+Bridge refresh is now wired to `session.end` and `tasks.complete` lifecycle hooks with a 30-second debounce to prevent thrashing during rapid completions.
+
+#### T139 — Context-Aware Bridge Generation
+
+`generateContextAwareContent` in `memory/memory-bridge.ts` uses `hybridSearch()` scoped to the current session and focused task to generate bridge content within the `brain.memoryBridge.maxTokens` budget. Activated when `brain.memoryBridge.contextAware: true`.
+
+#### T140 — Session Summarization
+
+`buildSummarizationPrompt` and `ingestStructuredSummary` in `memory/session-memory.ts` deliver dual-mode summarization:
+
+- **Prompt mode**: Returns a `memoryPrompt` string on `SessionEndResult` for agent-side summarization
+- **Ingestion mode**: Accepts a structured `SessionSummaryInput` via `sessionSummary` on `SessionEndParams` and auto-ingests to brain.db
+
+#### T141 — Auto-Link to Focused Task
+
+Observations written during a focused-task session are automatically linked to that task via `brain_memory_links`, enabling task-scoped brain retrieval.
+
+#### T142 — Embedding Backfill CLI
+
+`cleo backfill --embeddings` retroactively generates embeddings for existing brain entries with progress reporting. Dispatches to `runEmbeddingBackfill()` in the memory module.
+
+#### T143 — Brain Maintenance Command
+
+`cleo brain maintenance` runs combined maintenance in order: temporal decay → consolidation → embedding backfill. Individual steps can be skipped via `--skip-decay`, `--skip-consolidation`, `--skip-embeddings`. The `runBrainMaintenance()` function is exported from the internal barrel.
+
+#### T144 — Cross-Provider Transcript Extraction
+
+`AdapterHookProvider` gains an optional `getTranscript()` method. The Claude Code adapter implements this to expose session transcript content. `extractFromTranscript()` in `memory/auto-extract.ts` processes the transcript and emits observation candidates.
+
+#### T145 — Updated Injection Templates
+
+`CLEO-INJECTION.md` templates updated with a Memory Automation section documenting `brain.embedding`, `brain.memoryBridge`, and `brain.summarization` config keys for agent awareness.
+
+#### T146 — CLEO-BRAIN-SPECIFICATION.md v2.0.0
+
+Updated brain specification to v2.0.0 reflecting the full automation layer: embedding pipeline, lifecycle hooks, summarization protocol, and maintenance command.
+
+#### New Runtime Dependency
+
+| Package | Version | Load Strategy |
+|---------|---------|---------------|
+| `@xenova/transformers` | `^2.17.2` | Dynamic import (only when `brain.embedding.enabled: true`) |
+
 ---
 
 ## 16. Build Architecture