@mubit-ai/ai-sdk

Vercel AI SDK middleware backed by MuBit memory engine.

Wraps any AI SDK model with automatic memory injection and interaction capture. Targets the current AI SDK v3 language-model middleware contract (LanguageModelV3Middleware, specificationVersion: "v3") and works with ai@^6. The peer dependency is ai >=5; the legacy v2 contract is deprecated.

Install

npm install @mubit-ai/ai-sdk ai

Usage

import { wrapLanguageModel } from "ai";
import { mubitMemoryMiddleware } from "@mubit-ai/ai-sdk";

const middleware = mubitMemoryMiddleware({
  apiKey: "mbt_...",
  sessionId: "session-1",
});

const model = wrapLanguageModel({
  model: openai("gpt-4o"),
  middleware,
});

// generateText AND streamText calls through this model
// automatically get lessons injected + interactions captured.
const { text } = await generateText({ model, prompt: "Hello" });

// Close the v0.7.0 attribution loop once you know the outcome.
await middleware.recordOutcome({
  outcome: "success",
  entry_ids: middleware.getLastReferenceIds(), // credit recalled evidence
  verified_in_production: true,
});

How it works

The middleware has three hooks:

1. transformParams — Before each LLM call (generate or stream), fetches relevant context + reinforced lessons from MuBit and injects them into the system prompt. Evidence reference_ids are surfaced via getLastReferenceIds().
2. wrapGenerate — After each generateText call, captures the interaction and ingests it to MuBit (fire-and-forget, with an idempotency key).
3. wrapStream — Tees the streamText output so chunks flow through untouched, then ingests the completed interaction once the stream finishes (fire-and-forget).

Attribution loop (v0.7.0)

The returned middleware object exposes attribution helpers, also available as standalone exports (recordOutcome, recordStepOutcome):

• getLastReferenceIds() — reference_ids of the most recently surfaced evidence, for crediting specific entries.
• recordOutcome({ reference_id?, outcome, entry_ids?, verified_in_production?, signal?, rationale? }) — record a run-level (or entry-level) reinforcement outcome.
• recordStepOutcome({ step_id, step_name?, outcome?, signal?, rationale?, directive_hint? }) — record a step-level process-reward outcome.

All writes (remember/ingest) thread an idempotency_key so retried or double-captured interactions dedup server-side.

Options

| Option | Default | Description | | --- | --- | --- | | apiKey | MUBIT_API_KEY | MuBit API key | | endpoint | MUBIT_ENDPOINT / http://127.0.0.1:3000 | MuBit endpoint URL | | sessionId | "default" | Session/run ID for scoping memory | | userId | "" | User ID for scoping memory | | agentId | "ai-sdk" | Agent ID for attribution | | maxTokenBudget | 2048 | Token budget for context retrieval | | entryTypes | server default | Memory entry types to retrieve | | sections | server default | Context sections to include | | injectLessons | true | Whether to inject lessons before calls | | captureInteractions | true | Whether to capture interactions after calls | | failOpen | true | Proceed without lessons on error | | mubitClient | auto-created | Pre-configured MuBit client |

Features

• TTL-bounded in-memory cache (30s, 100 entries) for context retrieval
• Fail-open by default — LLM calls proceed even if MuBit is unreachable
• Fire-and-forget ingestion — doesn't block on capture
• Idempotency-keyed writes — retried/streamed captures dedup server-side
• Built on the public MuBit SDK surface only (no private internals)
• Compatible with generateText, streamText, and any AI SDK v3 model

Development

node --test test/**/*.test.js

License

Apache-2.0