fleet-memory/hindsight-integrations/ai-sdk
Nicolò Boschi 7e339e1677
feat: ai sdk integration (#299)
* feat: ai sdk integration

* more fixes

* fix(security): mental model refresh tag-based security

- Mental model refresh now passes tags with all_strict matching
- Consolidation only triggers refresh for mental models with matching tags
- Consolidation filters related observations by tags (all_strict)
- Added tests to verify tag-based security boundaries
- Updated OpenAPI spec to include tags and text_preview in list_documents
- Added tags column to documents UI table

* chore: regenerate OpenAPI spec after rebase

* fix: improve consolidation prompt for contradiction handling and mental model refresh security

- Enhanced consolidation prompt to be more explicit about capturing temporal changes in contradictions
- Fixed mental model refresh security: tagged memories now only trigger refresh of mental models with matching tags
- Added stricter tag filtering to prevent cross-scope mental model refreshes

Fixes test_consolidation_merges_contradictions by improving LLM instructions to use temporal markers like "used to X, now Y" when merging contradictory facts.

Note: test_refresh_with_tags_only_accesses_same_tagged_models still needs investigation - REFLECT operation may need additional tag filtering.

* fix: mental model refresh security - proper tag filtering in search

Fixed tool_search_mental_models to properly handle all_strict tag matching mode by using the centralized build_tags_where_clause function. Previously, the function only handled "all" vs "any" modes and always included untagged mental models when using non-"all" modes.

This ensures that when a tagged mental model is refreshed with all_strict matching, it cannot access untagged mental models, preventing cross-scope information leakage.

Fixes test_refresh_with_tags_only_accesses_same_tagged_models.

Note: test_sensory_dimension_preservation is failing but this is a pre-existing issue on main branch - the LLM model (gpt-oss-20b) is not extracting facts from sensory text. Not related to security changes.

* chore: apply formatting from pre-commit hook

* fix: allow untagged mental models to be refreshed by any consolidation

Untagged mental models are considered "global" and should be refreshed
by any consolidation, regardless of whether tagged or untagged memories
were consolidated. This maintains security boundaries while allowing
global mental models to stay fresh.

When tagged memories are consolidated:
- Refresh mental models with matching tags (security boundary)
- Also refresh untagged mental models (they're global)
- DO NOT refresh mental models with different tags

When untagged memories are consolidated:
- Only refresh untagged mental models
- DO NOT refresh tagged mental models (security boundary)

Fixes test_consolidation_only_refreshes_matching_tagged_models.
2026-02-04 20:25:59 +01:00
..
src feat: ai sdk integration (#299) 2026-02-04 20:25:59 +01:00
.gitignore feat: ai sdk integration (#299) 2026-02-04 20:25:59 +01:00
package-lock.json feat: ai sdk integration (#299) 2026-02-04 20:25:59 +01:00
package.json feat: ai sdk integration (#299) 2026-02-04 20:25:59 +01:00
README.md feat: ai sdk integration (#299) 2026-02-04 20:25:59 +01:00
tsconfig.json feat: ai sdk integration (#299) 2026-02-04 20:25:59 +01:00
vitest.config.ts feat: ai sdk integration (#299) 2026-02-04 20:25:59 +01:00

Hindsight Memory Integration for Vercel AI SDK

Give your AI agents persistent, human-like memory using Hindsight with the Vercel AI SDK.

Features

  • Three Memory Operations: retain (store), recall (retrieve), and reflect (reason over memories)
  • Multi-User Support: Dynamic bank IDs per call for multi-user/multi-tenant scenarios
  • Full API Coverage: Complete parameter support for all Hindsight operations
  • Type-Safe: Full TypeScript support with Zod schemas for validation
  • AI SDK 6 Native: Works seamlessly with generateText, streamText, and ToolLoopAgent

Installation

npm install @vectorize-io/hindsight-ai-sdk ai zod

You'll also need a Hindsight client. Choose one:

Option A: TypeScript/JavaScript Client

npm install @vectorize-io/hindsight-client

Option B: Direct HTTP Client (no additional dependencies)

// See "HTTP Client Example" below

Quick Start

1. Set up your Hindsight client

import { HindsightClient } from '@vectorize-io/hindsight-client';

const hindsightClient = new HindsightClient({
  apiUrl: process.env.HINDSIGHT_API_URL || 'http://localhost:8000',
});

2. Create Hindsight tools

import { createHindsightTools } from '@vectorize-io/hindsight-ai-sdk';

const tools = createHindsightTools({
  client: hindsightClient,
});

3. Use with AI SDK

import { generateText } from 'ai';
import { anthropic } from '@ai-sdk/anthropic';

const result = await generateText({
  model: anthropic('claude-sonnet-4-20250514'),
  tools,
  prompt: 'Remember that Alice loves hiking and prefers spicy food',
});

console.log(result.text);

Full Example: Memory-Enabled Chatbot

import { HindsightClient } from '@vectorize-io/hindsight-client';
import { createHindsightTools } from '@vectorize-io/hindsight-ai-sdk';
import { streamText } from 'ai';
import { anthropic } from '@ai-sdk/anthropic';

// Initialize Hindsight
const hindsightClient = new HindsightClient({
  apiUrl: 'http://localhost:8000',
});

const tools = createHindsightTools({ client: hindsightClient });

// Chat with memory
const result = await streamText({
  model: anthropic('claude-sonnet-4-20250514'),
  tools,
  system: `You are a helpful assistant with long-term memory.

IMPORTANT:
- Before answering questions, use the 'recall' tool to check for relevant memories
- When users share important information, use the 'retain' tool to remember it
- For complex questions requiring synthesis, use the 'reflect' tool
- Always pass the user's ID as the bankId parameter

Your memory persists across sessions!`,
  prompt: 'Remember that I am Alice and I love hiking',
});

for await (const chunk of result.textStream) {
  process.stdout.write(chunk);
}

API Reference

createHindsightTools(options)

Creates AI SDK tool definitions for Hindsight memory operations.

Parameters:

  • options.client: HindsightClient - Hindsight client instance
  • options.retainDescription: string (optional) - Custom description for the retain tool
  • options.recallDescription: string (optional) - Custom description for the recall tool
  • options.reflectDescription: string (optional) - Custom description for the reflect tool

Returns: Object with three tools: retain, recall, and reflect

Tool: retain

Store information in long-term memory.

Parameters:

  • bankId: string - Memory bank ID (usually the user ID)
  • content: string - Content to store
  • documentId: string (optional) - Document ID for grouping/upserting
  • timestamp: string (optional) - ISO timestamp for when the memory occurred
  • context: string (optional) - Additional context about the memory

Returns:

{
  success: boolean;
  itemsCount: number;
}

Tool: recall

Search memory for relevant information.

Parameters:

  • bankId: string - Memory bank ID
  • query: string - What to search for
  • types: string[] (optional) - Filter by fact types
  • maxTokens: number (optional) - Maximum tokens to return
  • budget: 'low' | 'mid' | 'high' (optional) - Processing budget
  • queryTimestamp: string (optional) - Query from a specific time (ISO format)
  • includeEntities: boolean (optional) - Include entity observations
  • includeChunks: boolean (optional) - Include raw chunks

Returns:

{
  results: Array<{
    id: string;
    text: string;
    type?: string;
    entities?: string[];
    context?: string;
    occurred_start?: string;
    occurred_end?: string;
    mentioned_at?: string;
    document_id?: string;
    metadata?: Record<string, string>;
    chunk_id?: string;
  }>;
  entities?: Record<string, EntityState>;
}

Tool: reflect

Analyze memories to form insights and generate contextual answers.

Parameters:

  • bankId: string - Memory bank ID
  • query: string - Question to reflect on
  • context: string (optional) - Additional context for reflection
  • budget: 'low' | 'mid' | 'high' (optional) - Processing budget

Returns:

{
  text: string;
  basedOn?: Array<{
    id?: string;
    text: string;
    type?: string;
    context?: string;
    occurred_start?: string;
    occurred_end?: string;
  }>;
}

Advanced Usage

Custom Tool Descriptions

Customize tool descriptions to guide model behavior:

const tools = createHindsightTools({
  client: hindsightClient,
  retainDescription: 'Store user preferences and important facts. Always include context.',
  recallDescription: 'Search past conversations. Use specific queries for best results.',
  reflectDescription: 'Synthesize insights from memories. Use for complex questions.',
});

Multi-User Scenarios

Each tool call accepts a bankId parameter, making it easy to support multiple users:

const result = await generateText({
  model: anthropic('claude-sonnet-4-20250514'),
  tools,
  prompt: `User ID: ${userId}\n\nRemember that I prefer dark mode`,
});

The model will automatically pass the user ID to the tools.

Using with ToolLoopAgent

import { ToolLoopAgent, stopWhen, stepCountIs } from 'ai';

const agent = new ToolLoopAgent({
  model: anthropic('claude-sonnet-4-20250514'),
  tools,
  instructions: `You are a personal assistant with long-term memory.

    Always check memory before responding using the recall tool.
    Store important user preferences with the retain tool.
    Use the reflect tool to analyze patterns in the user's behavior.`,
  stopWhen: stepCountIs(10),
});

const result = await agent.generate({
  prompt: 'What did I say I wanted to work on this week?',
});

HTTP Client Example

If you prefer not to install the full Hindsight client, you can use a simple HTTP client:

import type { HindsightClient } from '@vectorize-io/hindsight-ai-sdk';

const httpClient: HindsightClient = {
  async retain(bankId, content, options = {}) {
    const response = await fetch(`${HINDSIGHT_URL}/v1/default/banks/${bankId}/memories/retain`, {
      method: 'POST',
      headers: { 'Content-Type': 'application/json' },
      body: JSON.stringify({
        content,
        timestamp: options.timestamp,
        context: options.context,
        metadata: options.metadata,
        document_id: options.documentId,
        async: options.async,
      }),
    });
    return response.json();
  },

  async recall(bankId, query, options = {}) {
    const response = await fetch(`${HINDSIGHT_URL}/v1/default/banks/${bankId}/memories/recall`, {
      method: 'POST',
      headers: { 'Content-Type': 'application/json' },
      body: JSON.stringify({
        query,
        types: options.types,
        max_tokens: options.maxTokens,
        budget: options.budget,
        trace: options.trace,
        query_timestamp: options.queryTimestamp,
        include_entities: options.includeEntities,
        max_entity_tokens: options.maxEntityTokens,
        include_chunks: options.includeChunks,
        max_chunk_tokens: options.maxChunkTokens,
      }),
    });
    return response.json();
  },

  async reflect(bankId, query, options = {}) {
    const response = await fetch(`${HINDSIGHT_URL}/v1/default/banks/${bankId}/reflect`, {
      method: 'POST',
      headers: { 'Content-Type': 'application/json' },
      body: JSON.stringify({
        query,
        context: options.context,
        budget: options.budget,
      }),
    });
    return response.json();
  },
};

const tools = createHindsightTools({ client: httpClient });

Running Hindsight Locally

The easiest way to run Hindsight for development:

# Install and run with embedded mode (no setup required)
uvx hindsight-embed@latest -p myapp daemon start

# The API will be available at http://localhost:8000

For production deployments, see the Hindsight Documentation.

TypeScript Types

All types are exported for your convenience:

import type {
  Budget,
  HindsightClient,
  HindsightTools,
  HindsightToolsOptions,
  RecallResult,
  RecallResponse,
  ReflectFact,
  ReflectResponse,
  RetainResponse,
  EntityState,
  ChunkData,
} from '@vectorize-io/hindsight-ai-sdk';

Documentation & Resources

License

MIT

Support

For issues and questions: