anansi-ai-sdk

Vercel AI SDK middleware for Anansi memory. Wrap any language model — before each request the user's synthesized profile is injected as a system message; optionally each user turn is auto-ingested back into memory.

import { withAnansi } from 'anansi-ai-sdk';
import { openai } from '@ai-sdk/openai';
import { generateText } from 'ai';

const model = withAnansi(openai('gpt-4o'), {
  apiKey: 'ans_...',
  userId: 'user_123',
  mode: 'full',       // 'profile' | 'query' | 'full'
  ingestAfter: true,  // auto-ingest conversation turns
});

const result = await generateText({ model, prompt: 'What stack should I use?' });

Modes

| mode | behavior | |---|---| | profile | inject the synthesized static + dynamic profile only (no search) | | query | search memory with the latest user message and inject relevant[] only | | full | both (default) |

Options

• apiKey (required) — Anansi API key
• userId (required) — memory is scoped per user
• mode — see above, default full
• ingestAfter — fire-and-forget ingest of each user message, default false
• sessionId — tag ingested turns for session-scoped retrieval
• baseUrl — defaults to the hosted Anansi API

Works with both LanguageModelV1 and LanguageModelV2 models — the wrapper intercepts doGenerate/doStream structurally, so ai is only an optional peer dependency.