Changelog.

Product updates across supermemory API, MCP, plugins, SDKs, local tooling, and research.

MCP retries instead of signing you out during an auth outage#

A brief authentication outage used to return invalid_token, the protocol signal to discard your credential, so every connected MCP client was logged out. Those requests now return 503 with Retry-After so clients wait and retry.

Improvements

  • withSupermemory in @supermemory/tools now takes an apiKey option instead of only reading the environment variable, so it works with secrets managers, edge runtimes, and per-request keys.
  • The ai-sdk package re-exports PromptTemplate, MemoryPromptData, and WithSupermemoryOptions, so the documented custom-template example compiles.

Fixes

  • Search and graph tools now bound their page and limit inputs, and whoAmI no longer returns the transport session id.

withSupermemory now takes the key directly, so it works with a secrets manager, an edge runtime, or a different key per request:

import { withSupermemory } from "@supermemory/tools/ai-sdk"

const model = withSupermemory(openai("gpt-5.6"), "user-123", {
  apiKey: process.env.MY_VAULT_KEY,
})

Requires @supermemory/tools 2.2.0 or later:

npm install @supermemory/tools@latest

Backfill historical documents with documentDate on ingest#

Set documentDate on each document so historical content lands on the date it actually happened instead of the date you uploaded it, which is what makes time-based search work on backfilled data.

Improvements

  • A new guide walks through backfilling dated documents through the batch ingestion API, with TypeScript and Python examples.

Fixes

  • Fixed an issue where SDK responses dropped fields the API returned, across 15 endpoints.

Send up to 600 documents per request to POST /v3/documents/batch, sorted oldest to newest:

{
  "documents": [
    {
      "id": "msg-2019-04-02-118",
      "content": "Decided to move the billing cutover to Q3.",
      "documentDate": "2019-04-02T09:15:00Z"
    }
  ]
}

Without documentDate a backfilled document is dated the day you uploaded it, so a question like "what did we decide in 2019" will not find it.

Nova message editing and prompt queue, Claude Code /supermemory:status, and Cursor session memory#

Nova adds message editing, a three-prompt queue, and Instant and Thinking modes; Claude Code gains /supermemory:status for authentication, project, and API key details; Cursor loads project memory at session start and saves it when the session ends.

Codex plugin routes memories into project Spaces#

Codex can organize captured work into custom memory containers and keep project Spaces named from the context it sees.

Fixes

  • Fixed the Python OpenAI SDK injecting an empty system prompt when no memories were found.

Persistent memory plugin for OpenAI Codex#

Install the Codex plugin to recall project and user memory, capture completed sessions, and explicitly search or save context.

Fixes

  • The AI SDK wrapper now continues without memory when retrieval fails or exceeds five seconds.

MemoryBench providers, Claude skills, and container tag filtering#

MemoryBench gains a Claude skill for scaffolding provider adapters and running benchmarks, new filesystem-memory and hybrid-RAG providers for comparison, and the Supermemory repository gets a Claude skill with setup guidance and TypeScript and Python examples.

AI SDK strict mode, MCP recall, and browser extension fixes#

AI SDK memory tools support strict schema validation for models that require every parameter, MCP recall results include the memory text next to each match percentage, and pages saved by the browser extension keep their original source link.

YOUR PRIVACY

Change or withdraw any time via Cookie settings in the footer. Read our cookie notice.