wolbarg 0.5.0 → 0.5.2

package/CHANGELOG.md

@@ -5,6 +5,31 @@ All notable changes to this project are documented here.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/).
 
+## [0.5.2] — 2026-07-19
+
+### Added
+
+- **`rememberFromMessages()` (experimental)** — conversation → memory bridge with `mode: "raw"` (default, no LLM) and optional `mode: "extract"` via configured `llm`
+- **Vercel AI SDK adapter example** — `examples/adapters/vercel-ai/` (now uses official `@wolbarg/vercel-ai` middleware)
+- **Companion package `@wolbarg/vercel-ai`** — Language Model Middleware (`wolbargMiddleware` + `wrapLanguageModel`) for automatic recall / remember (published separately under `packages/vercel-ai/`)
+
+### Compatibility
+
+- Additive only. Omit the new method and behavior matches **0.5.1**. Experimental API may change before 1.0 — pin versions if you depend on it.
+- Core `wolbarg` remains framework-agnostic; AI SDK types live only in `@wolbarg/vercel-ai`.
+- `@wolbarg/vercel-ai@1` requires **AI SDK v7+** (`ai@^7`). Upgrade from AI SDK v4 before adopting the middleware.
+
+## [0.5.1] — 2026-07-19
+
+
+### Fixed
+
+- **SQLite vec0 rowid binding on Linux** — bind `memory_rowid` as `BigInt` for `node:sqlite` + sqlite-vec so CI / Linux inserts and KNN search do not fail on integer PK binds
+
+### Compatibility
+
+- Drop-in patch for **0.5.0**. No API or schema changes.
+
 ## [0.5.0] — 2026-07-19
 
 ### Added
@@ -166,6 +191,7 @@ const ctx = wolbarg({
 
 - Initial npm release path (pre–modular storage / ingest)
 
+[0.5.1]: https://github.com/wolbarg/wolbarg/releases/tag/v0.5.1
 [0.5.0]: https://github.com/wolbarg/wolbarg/releases/tag/v0.5.0
 [0.4.0]: https://github.com/wolbarg/wolbarg/releases/tag/v0.4.0
 [0.3.2]: https://github.com/wolbarg/wolbarg/releases/tag/v0.3.2

package/README.md

@@ -1,59 +1,134 @@
-<p align="center">
-  <img src="./assets/wolbarg-banner.png" alt="Wolbarg Logo: Memory Infrastructure for AI Agents" width="720" />
-</p>
+![Wolbarg — Memory Infrastructure for AI Agents](./assets/wolbarg-banner.png)
 
 # Wolbarg
 
-**Modular, provider-agnostic semantic memory for AI agents (v0.5.0).**
+**Modular, provider-agnostic semantic memory for AI agents.**
 
 [![npm version](https://img.shields.io/npm/v/wolbarg.svg)](https://www.npmjs.com/package/wolbarg)
+[![SDK CI](https://github.com/wolbarg/wolbarg/actions/workflows/sdk-ci.yml/badge.svg?branch=main)](https://github.com/wolbarg/wolbarg/actions/workflows/sdk-ci.yml)
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](./LICENSE)
+[![Node.js](https://img.shields.io/badge/node-%3E%3D22.5-brightgreen.svg)](https://nodejs.org)
 [![Docs](https://img.shields.io/badge/docs-wolbarg.com-black)](https://wolbarg.com)
-[![Benchmarks](https://img.shields.io/badge/benchmarks-Storage%20suite-black)](https://wolbarg.com/benchmarks)
+[![Benchmarks](https://img.shields.io/badge/benchmarks-v0.4%20stress-black)](https://wolbarg.com/benchmarks)
 
-## What's new in 0.5.0
+```bash
+npm install wolbarg
+```
+
+Wolbarg is **memory infrastructure**, not an agent framework. Agents call `remember()` / `recall()` (and optionally ingest, compress, subscribe, and link memories in a graph). You bring SQLite or PostgreSQL, any OpenAI-compatible embedding API, and optional peers for PDF/DOCX/OCR/Neo4j.
 
-Optional **graph memory** (`sqliteGraph` / `neo4jGraph`), `linkMemories` / `getRelated`, `recall({ includeGraph: true })`, cascade forget, graph-aware checkpoints, and **Wolbarg Studio** graph canvas.
+**Current release: [v0.5.2](./CHANGELOG.md)** — experimental `rememberFromMessages()` (chat → memory) and a [Vercel AI SDK adapter example](../examples/adapters/vercel-ai/). Still includes 0.5 graph memory (SQLite + Neo4j), `includeGraph` recall, and [Wolbarg Studio](https://wolbarg.com/docs/observability).
 
-Full notes: [CHANGELOG](./CHANGELOG.md) · [Docs — What's New](https://wolbarg.com/docs/guides/whats-new) · [Graph memory](https://wolbarg.com/docs/graph-memory) · [Studio](https://wolbarg.com/docs/observability)
+---
 
-## Benchmarks
+## Why Wolbarg?
+
+| Problem | What Wolbarg does |
+| --- | --- |
+| Agents forget between sessions | Durable semantic memory on **your** disk or Postgres |
+| Vendor lock-in | Pluggable embeddings, LLM, rerankers, OCR, vision, storage, graph |
+| Vectors alone are not enough | Hybrid search (semantic + BM25), metadata filters, MMR, optional rerank |
+| Multi-agent writes collide | SQLite `BEGIN IMMEDIATE` + retries; Postgres row-level locking |
+| Duplicate facts pile up | Opt-in write-time dedupe / upsert |
+| Repeated embeds cost money | Transparent embedding cache (default on) |
+| Need structure, not only similarity | Optional **graph** — `linkMemories` / `getRelated` (SQLite ↔ Neo4j) |
+| Hard to debug memory | Telemetry + **Wolbarg Studio** (local, read-only) |
+
+---
+
+## Features
+
+- **Storage** — SQLite (`node:sqlite` + sqlite-vec) or PostgreSQL (+ optional pgvector)
+- **Recall** — semantic, hybrid, metadata filters, MMR, rerank, `explain: true`
+- **Ingest** — text / markdown / PDF / DOCX / images (peers for PDF/DOCX/OCR)
+- **Compress** — optional LLM summarization with archive lineage
+- **`subscribe()`** — real-time change events (SQLite in-process; Postgres LISTEN/NOTIFY)
+- **Concurrency** — multi-writer SQLite hardening (`WOLBARG_STORAGE_LOCKED`)
+- **Embedding cache** — `hash(content) + model`
+- **Dedupe / upsert** — opt-in exact / near / exact-or-near
+- **Graph memory (0.5)** — `sqliteGraph` / `neo4jGraph`, `linkMemories`, `getRelated`, `includeGraph`
+- **Conversation bridge (0.5.2, experimental)** — `rememberFromMessages()` raw or LLM extract
+- **Framework adapter** — official [`@wolbarg/vercel-ai`](../packages/vercel-ai/) middleware (`wrapLanguageModel`)
+- **Checkpoints / export** — file-backed SQLite snapshots (+ graph file when applicable)
+- **Telemetry** — independent event DB; Studio dashboard, Trace Explorer, graph canvas
+
+Docs: [Getting started](https://wolbarg.com/docs/getting-started) · [Vercel AI](https://wolbarg.com/docs/integrations/vercel-ai) · [rememberFromMessages](https://wolbarg.com/docs/api/remember-from-messages) · [Graph memory](https://wolbarg.com/docs/graph-memory) · [Studio](https://wolbarg.com/docs/observability)
+
+---
+
+## Benchmarks (v0.4.0 stress · mock embeddings)
+
+Published dual-backend **v4-stress** suite (2026-07-18 · Node v24.13.1 · win32/arm64 · 8 CPUs). Mock embeddings isolate SDK + database cost — not LIVE provider latency.
+
+### Headlines
+
+| Metric | SQLite | PostgreSQL |
+| --- | --- | --- |
+| Cold `ready()` | **16.18 ms** | **91.39 ms** |
+| Batch remember (200) | **5,795 ops/s** | **2,795 ops/s** |
+| Bulk insert 2k | **7,509 ops/s** | **4,085 ops/s** |
+| Recall p95 @ 2k | **4.83 ms** | **141.5 ms** |
+| 16 writers throughput | **8,660 ops/s** | **3,335 ops/s** |
+| Embedding cache speedup | **1.47×** | **1.18×** |
+| Suite result | 25 pass / 0 fail | 21 pass / 0 fail / 4 skip\* |
+
+\*Postgres skips SQLite-only paths (telemetry EventDatabase, file checkpoints, export/import).
+
+### SQLite — stress & concurrency
+
+| Case | Result |
+| --- | --- |
+| Bulk insert 2k + search | 7,509 insert ops/s · recall p50 **4.12 ms** · p95 **4.83 ms** |
+| 8 writers × 20 ops | **6,084** ops/s · p95 **3.05 ms** · 0 failures |
+| 16 writers × 20 ops | **8,660** ops/s · p95 **2.46 ms** · 0 failures |
+| 32 writers × 20 ops | **6,798** ops/s · p95 **22.94 ms** · 0 failures |
+| Mixed read/write storm | 0 failures |
 
-**v0.2.1 Storage suite** (mock embeddings · scale quick) — dual-backend SQLite + PostgreSQL:
+### PostgreSQL — stress & concurrency
 
-| Metric | Result |
+| Case | Result |
 | --- | --- |
-| SQLite search @ 1k | **2.02 ms** |
-| SQLite insert @ 1k | **1.72k ops/sec** |
-| SQLite cold start | **7.91 ms** |
-| Postgres 16 writers | **1.12k ops/sec** |
+| Bulk insert 2k + search | 4,085 insert ops/s · recall p50 **23.29 ms** · p95 **141.5 ms** |
+| 8 writers × 20 ops | **2,555** ops/s · p95 **8.51 ms** · 0 failures |
+| 16 writers × 20 ops | **3,335** ops/s · p95 **9.48 ms** · 0 failures |
+| 32 writers × 20 ops | **3,802** ops/s · p95 **14.77 ms** · 0 failures |
+| Mixed read/write storm | 0 failures |
 
-Storage (mock) ≠ LIVE (real providers). Full page: [wolbarg.com/benchmarks](https://wolbarg.com/benchmarks) · [methodology](https://wolbarg.com/docs/benchmarks) · [raw JSON](https://wolbarg.com/benchmarks/benchmark.json).
+**Artifacts:** [SQLite JSON](https://wolbarg.com/benchmarks/version-0.4.0-sqlite-benchmark.json) · [SQLite MD](https://wolbarg.com/benchmarks/version-0.4.0-sqlite-benchmark.md) · [Postgres JSON](https://wolbarg.com/benchmarks/version-0.4.0-postgres-benchmark.json) · [Postgres MD](https://wolbarg.com/benchmarks/version-0.4.0-postgres-benchmark.md) · [Interactive page](https://wolbarg.com/benchmarks) · [Methodology](https://wolbarg.com/docs/benchmarks)
+
+Also published: [embedding-cache](https://wolbarg.com/benchmarks/embedding-cache.json) · [multiprocess concurrency](https://wolbarg.com/benchmarks/multiprocess-concurrency.json)
+
+---
 
 ## Installation
 
 ```bash
 npm install wolbarg
+# or: pnpm add wolbarg · yarn add wolbarg · bun add wolbarg
 ```
 
-Node.js **22.5+**.
+**Requires Node.js 22.5+** (built-in `node:sqlite`).
 
-### Optional peers (install when you use the feature)
+### Optional peers
+
+Install only what you use:
 
 | Peer | Required for |
 | --- | --- |
-| `pg` | `postgres({ … })` storage |
-| `neo4j-driver` | `neo4jGraph({ … })` graph backend |
-| `pdf-parse` (pin `@1.1.4`) | `ingest()` of `.pdf` files |
-| `mammoth` | `ingest()` of `.docx` files |
-| `tesseract.js` | OCR provider for images |
+| `pg` | PostgreSQL storage |
+| `neo4j-driver` | `neo4jGraph(...)` |
+| `pdf-parse@1.1.4` | PDF ingest |
+| `mammoth` | DOCX ingest |
+| `tesseract.js` | OCR |
 
 ```bash
-# Example: PDF + DOCX ingest
-npm install pdf-parse@1.1.4 mammoth
+npm install pg neo4j-driver          # production storage + graph
+npm install pdf-parse@1.1.4 mammoth  # document ingest
 ```
 
-**Important:** these packages are **not** bundled with `Wolbarg`. If you call `ingest()` on PDF/DOCX without the matching peer, Wolbarg throws a configuration error at use time (not at import). Plain `.txt` / `.md` / `.csv` / `.json` need no extras.
+Peers are **not** bundled. Missing peers fail at use time for that path — not at import. Plain `.txt` / `.md` / `.csv` / `.json` and SQLite graph need no extras.
+
+---
 
 ## Quick start
 
@@ -72,66 +147,83 @@ const ctx = wolbarg({
     model: "gpt-4.1-mini",
   }),
   keywordSearch: bm25(),
+  concurrency: { maxRetries: 5 },
+  embeddingCache: { enabled: true },
+  memory: { dedupe: { enabled: true, strategy: "exact-or-near" } },
   telemetry: {
     enabled: true,
     database: { provider: "sqlite", url: "./telemetry.db" },
     level: "debug",
-    captureQueries: true,
-    captureLatency: true,
   },
 });
 
-await ctx.remember({
+await ctx.ready();
+
+const saved = await ctx.remember({
   agent: "research",
   content: { text: "Stripe supports recurring invoices." },
   metadata: { topic: "billing" },
 });
+// saved.action === "created" | "updated"
+
+const hits = await ctx.recall({
+  query: "How do recurring invoices work?",
+  topK: 5,
+  hybrid: true,
+});
+
+ctx.subscribe({ organization: "my-org" }, (e) => {
+  console.log(e.event, e.memoryId);
+});
 
-const hits = await ctx.recall({ query: "invoices", topK: 5 });
-const explained = await ctx.recall({ query: "invoices", explain: true });
 await ctx.checkpoint("before-compress");
+await ctx.close();
 ```
 
-### Optional graph (0.5)
+### Graph memory (0.5)
+
+Same typed API locally and in production — swap only the factory:
 
 ```ts
-import { wolbarg, openaiEmbedding, sqliteGraph } from "wolbarg";
+import { wolbarg, openaiEmbedding, sqliteGraph, neo4jGraph } from "wolbarg";
 
-const ctx = wolbarg({
-  organization: "my-org",
-  database: { provider: "sqlite", url: "./memory.db" },
-  embedding: openaiEmbedding({
-    apiKey: process.env.OPENAI_API_KEY!,
-    model: "text-embedding-3-small",
-  }),
-  graph: sqliteGraph({ path: "./graph.db" }),
-});
+// Local
+graph: sqliteGraph({ path: "./graph.db" })
 
+// Production
+graph: neo4jGraph({
+  url: process.env.NEO4J_URL!,
+  username: process.env.NEO4J_USER!,
+  password: process.env.NEO4J_PASSWORD!,
+})
+```
+
+```ts
 const a = await ctx.remember({
-  agent: "research",
+  agent: "support",
   content: { text: "Refunds take 5 business days." },
 });
 const b = await ctx.remember({
-  agent: "research",
+  agent: "support",
   content: { text: "Chargebacks escalate to risk." },
 });
+
 await ctx.linkMemories(a.id, b.id, "related_to");
-const related = await ctx.getRelated(a.id);
+const related = await ctx.getRelated(a.id, { depth: 1 });
+
+const withGraph = await ctx.recall({
+  query: "refunds",
+  includeGraph: true,
+});
+// withGraph[0].related — neighbors from the graph
 ```
 
-`new Wolbarg({ storage: sqlite("./memory.db"), embedding })` remains supported for backwards compatibility.
+Guide: [Graph memory](https://wolbarg.com/docs/graph-memory) · Prefer a [provider-isolated project layout](https://wolbarg.com/docs/installation#project-layout) so backend swaps stay one-file.
 
-## Quick start (legacy constructor)
+### Constructor DI (still supported)
 
 ```ts
-import {
-  Wolbarg,
-  sqlite,
-  openaiEmbedding,
-  openaiLlm,
-  bm25,
-  meta,
-} from "wolbarg";
+import { Wolbarg, sqlite, openaiEmbedding } from "wolbarg";
 
 const ctx = new Wolbarg({
   organization: "my-org",
@@ -140,78 +232,93 @@ const ctx = new Wolbarg({
     apiKey: process.env.OPENAI_API_KEY!,
     model: "text-embedding-3-small",
   }),
-  llm: openaiLlm({
-    apiKey: process.env.OPENAI_API_KEY!,
-    model: "gpt-4.1-mini",
-  }),
-  keywordSearch: bm25(),
-});
-
-await ctx.remember({
-  agent: "research",
-  content: { text: "Stripe supports recurring invoices." },
-  metadata: { topic: "billing" },
-});
-
-const hits = await ctx.recall({
-  query: "recurring invoices",
-  topK: 5,
-  hybrid: true,
-  filter: { metadata: meta.eq("topic", "billing") },
 });
 ```
 
 **Required:** `organization`, `storage` or `database`, `embedding`.  
-**Optional:** `llm`, `telemetry`, `keywordSearch`, `reranker`, `ocr`, `vision`, `chunking`, `compression`, `retrieval`, `checkpoint`.
+**Optional:** `llm`, `keywordSearch`, `reranker`, `ocr`, `vision`, `chunking`, `telemetry`, `concurrency`, `embeddingCache`, `memory.dedupe`, `graph`, `checkpoint` / `checkpointDirectory`.
 
-Calling `compress` without `llm` is a TypeScript error.
+---
 
-## API
+## API surface
 
-| Method | Description |
+| Method | Role |
 | --- | --- |
-| `remember` / `rememberBatch` | Store + embed |
-| `recall` / `recallBatch` | Semantic / hybrid search (`explain: true` for diagnostics) |
-| `ingest` | Documents → chunks → memories (**peers** for PDF/DOCX/OCR) |
-| `compress` | LLM summary (needs `llm`) |
-| `checkpoint` / `rollback` / `listCheckpoints` | First-party memory snapshots |
-| `export` / `import` | Portable SQLite memory bundles |
-| `forget` / `history` / `stats` / `clear` | Management |
-| `ready` / `close` | Lifecycle |
+| `remember` / `rememberBatch` | Embed + store (`RememberResult` includes `action`) |
+| `update` | Edit an existing memory by id |
+| `recall` / `recallBatch` | Semantic / hybrid search; `explain` · `includeGraph` |
+| `ingest` | Documents → chunks → memories |
+| `compress` | LLM summary (requires `llm`) |
+| `linkMemories` / `getRelated` | Graph edges / traversal (requires `graph`) |
+| `subscribe` | Real-time change callbacks |
+| `checkpoint` / `rollback` / `listCheckpoints` / … | SQLite snapshots (+ graph file when SQLite graph) |
+| `export` / `import` | Portable SQLite + manifest |
+| `forget` / `history` / `stats` / `clear` | Lifecycle (forget/clear cascade graph when configured) |
+| `ready` / `close` / `flushTelemetry` | Lifecycle |
+
+Full reference: [wolbarg.com/docs/api](https://wolbarg.com/docs/api)
+
+---
 
-## Studio
+## Wolbarg Studio
 
-Local telemetry dashboard (separate app — not bundled into the SDK):
+Local observability dashboard — **not** bundled in the npm package. The SDK writes telemetry; Studio reads it (and can open memory / graph files for the canvas).
 
 ```bash
-cd wolbarg_studio
+git clone https://github.com/Atharvmunde11/wolbarg-studio
+cd wolbarg-studio
 npm install
-npm run dev
+npm run dev   # http://localhost:3100
 ```
 
-Open http://localhost:3100 and connect to your `./telemetry.db`. Connections persist under the user config directory (`~/.wolbarg/studio.json` / AppData), never inside the project.
+Connect telemetry (`./telemetry.db`), optional memory DB, checkpoint directory, and graph backend. Screenshots and setup: [Observability & Studio](https://wolbarg.com/docs/observability).
 
-Telemetry SQLite files use an additive, versioned schema. On writable open, v1
-files are migrated in place to telemetry schema v2 and retain all existing
-events. V2 adds organization, agent, tags, checkpoint, persisted recall
-explanations, and measured stage spans. `SqliteEventDatabase.query()` can filter
-by `organization`, `agentId`, `tag`, and `checkpointId`; read-only access to an
-unmigrated v1 file remains supported.
+---
 
-Full documentation: [wolbarg.com](https://wolbarg.com/docs/introduction)
+## Documentation
 
-## Limitations (v0.3)
+| Topic | Link |
+| --- | --- |
+| Getting started | https://wolbarg.com/docs/getting-started |
+| Installation & project layout | https://wolbarg.com/docs/installation |
+| Configuration | https://wolbarg.com/docs/configuration |
+| Graph memory | https://wolbarg.com/docs/graph-memory |
+| Observability & Studio | https://wolbarg.com/docs/observability |
+| Concurrency | https://wolbarg.com/docs/concurrency |
+| Real-time events | https://wolbarg.com/docs/realtime-events |
+| Embedding cache | https://wolbarg.com/docs/embedding-cache |
+| Memory upsert | https://wolbarg.com/docs/memory-upsert |
+| What's new in 0.5 | https://wolbarg.com/docs/guides/whats-new |
+| Migration | https://wolbarg.com/docs/migration |
+| Limitations | https://wolbarg.com/docs/guides/limitations |
+| Changelog | [CHANGELOG.md](./CHANGELOG.md) |
+
+---
+
+## Limitations
+
+- **Node 22.5+** required (`node:sqlite`).
+- **Ingest peers** required for PDF / DOCX / OCR paths.
+- **PDF** text-layer via `pdf-parse`; scan PDFs need OCR/vision.
+- **Telemetry** is SQLite-only today (Postgres typed but not implemented).
+- **SQLite `subscribe()`** is in-process only; Postgres uses LISTEN/NOTIFY.
+- **Checkpoints / export** require file-backed SQLite memory (not `:memory:` / not Postgres).
+- **Neo4j** checkpoint / export / import throw `GraphCheckpointNotSupportedError` (refuse, don't skip).
+- **Cypher `query()`** is Neo4j-only; SQLite graph uses typed methods.
+- Not an agent framework, chat UI, or hosted vector SaaS.
+
+---
+
+## Upgrade
 
-- **Ingest peers are opt-in but required for those formats** — see table above.
-- **PDF text layer only** via `pdf-parse`; scan/image PDFs need OCR/vision (or a text PDF). Older pdf.js in `pdf-parse@1.1.4` may reject some modern PDFs.
-- **Node `node:sqlite` is experimental**; Node **22.5+** required.
-- **Postgres** memory storage needs `pg`; telemetry/checkpoints are SQLite-only in v0.3 (interfaces ready for Postgres).
-- **Not** an agent framework, chat UI, or hosted vector SaaS.
+```bash
+npm install wolbarg@^0.5.0
+```
 
-## Migration from 0.1 / 0.2
+0.5 is **additive**. Omitting `graph` keeps 0.4 behavior. See [Migration](https://wolbarg.com/docs/migration) and [CHANGELOG](./CHANGELOG.md).
 
-`init()` and `new Wolbarg({ storage })` still work. Prefer `wolbarg({ database, telemetry })`. LLM is optional. Schema auto-migrates to v2.
+---
 
 ## License
 
-MIT
+MIT © [Atharv Munde](https://github.com/Atharvmunde11)