supermemory

The withSupermemory function uses object spread to wrap the language model, but AI SDK 6 models store non-enumerable/prototype properties like specificationVersion, provider, modelId, and supportedUrls. Spread doesn't copy these, causing undefined values that trigger the error. Fix: replace the spread pattern in wrapVercelLanguageModel (packages/tools/src/vercel/index.ts) with the standard wrapLanguageModel middleware approach from 'ai' (import { wrapLanguageModel, LanguageModelV3Middleware } from 'ai'). As a temporary workaround, manually implement a wrapper using wrapLanguageModel until the library updates. Versions affected: @supermemory/tools@1.4.1, ai@6.0.158, @ai-sdk/provider@3.0.8.

**Cause:** Since supermemory v5.0.0-rc.6 (switch from Speakeasy to Fern codegen), the generated core contains literal dynamic imports — await import("fs") in generated/core/file/file.js and await import("stream") in generated/core/form-data-utils/FormDataWrapper.js. Bundlers like esbuild resolve literal import() specifiers at build time even if the code path never runs, so any browser/edge bundle (Cloudflare Workers, Vercel Edge, Convex default runtime, Vite/esbuild) fails with Could not resolve "fs" / Could not resolve "stream", even if you never upload a file.

**Workarounds:**
1. Pin to the last bundlable version: npm i supermemory@5.0.0-rc.5 (Speakeasy-generated builds bundle cleanly).
2. **Convex:** Move every module that imports the SDK into Convex's Node runtime by adding the 'use node' directive at the top of the file, since the default Convex runtime cannot use Node APIs.
3. **For SDK maintainers / local patching:** Apply a post-generation patch (e.g., in scripts/generate.ts) that replaces the literal specifiers with computed ones so esbuild leaves them as runtime imports: import(["f","s"].join("")) instead of import("fs") (note: const s = "fs"; import(s) and /* @vite-ignore */ comments do NOT work — esbuild with --minify-syntax inlines consts and still resolves them). Add /* webpackIgnore: true */ to also cover webpack. Node still resolves the computed form at runtime, so uploads keep working there.
4. Add a CI check that bundles the built package for a browser target (esbuild entry.js --bundle --platform=browser --format=esm) and fails on resolution errors to prevent regressions.

This is a known bug (v5.0.1) where the client.fetch() passthrough ignores the default environment. Unlike typed methods, it doesn't fall back to SupermemoryEnvironment.Default, so relative paths have no base URL to join against (throwing Failed to parse URL), and absolute URLs skip the auth check that attaches the Authorization header, causing 401s. Workaround: pass the base URL explicitly when creating the client:

import { Supermemory } from "supermemory"
import { SupermemoryEnvironment } from "supermemory/environments"

const client = new Supermemory({
  apiKey: "sm_test",
  baseUrl: SupermemoryEnvironment.Default, // or "https://api.supermemory.ai"
})

await client.fetch("/v3/settings") // now resolves and sends Bearer auth

Once baseUrl is set, both relative-path resolution and targetsBaseUrl auth attachment work correctly. A permanent fix requires resolving the base as this._options.baseUrl ?? this._options.environment ?? environments.SupermemoryEnvironment.Default in the generated client.

This happens because chunk boundaries are position-sensitive. An insertion or deletion near the beginning shifts every downstream boundary, so the content hashes no longer match and almost all chunks are re-embedded. Appends work because they do not shift existing boundaries.

Fix: upgrade the self-hosted server to a release containing PR fix(chunking): stable content-defined chunking with hash-based reuse planning (the fix for issue #1649). It uses stable/content-defined chunking and hash-based reuse so a mid-file edit invalidates only the affected chunks instead of the whole document.

Until you can upgrade, use these workarounds:
1. Prefer append-only updates when possible.
2. Split large living documents into smaller, relatively stable documents, such as one file per major Markdown heading.
3. Move frequently edited sections to the end of the document.
4. Use a faster embedding provider, GPU acceleration, or a smaller local embedding model to reduce the cost of unavoidable full re-embeds.

After upgrading, subsequent /v3/documents upserts with the same customId should reuse unchanged downstream chunks instead of re-embedding them.