flutter_agentic_memory 0.2.0

Semantic memory for Flutter AI agents — embeddings, on-device vector search, and long-term recall that plugs into flutter_agentic.

flutter_agentic_memory — Semantic Memory for Flutter AI Agents

Give your Flutter AI agent long-term memory. Embeddings (Gemini, OpenAI, or your own), on-device vector search with cosine similarity, persistent storage via Hive, and a drop-in SemanticMemory store for flutter_agentic agents that recalls past conversation by meaning, not just recency.

👉 View on pub.dev · API docs · GitHub

Part of the flutter_agentic family: flutter_agentic (core SDK) · flutter_agentic_graph · flutter_agentic_ui · flutter_agentic_tools

---

Installation

dependencies:
  flutter_agentic_memory: ^0.1.1

Quick start — agent that remembers

import 'package:flutter_agentic/flutter_agentic.dart';
import 'package:flutter_agentic_memory/flutter_agentic_memory.dart';

final memory = SemanticMemory(
  embedder: GeminiEmbeddingProvider(apiKey: 'YOUR_KEY'), // free tier works
  vectors: InMemoryVectorStore(),
);

final agent = AgenticAgent(
  provider: GeminiProvider(apiKey: 'YOUR_KEY'),
  memory: memory, // drop-in MemoryStore — nothing else changes
);

await agent.chat('Remember: I am vegetarian and allergic to peanuts.');
// ...hundreds of turns later:
final recalled = await memory.recall('default', 'what should I cook tonight?');
// → "Remember: I am vegetarian and allergic to peanuts." (score 0.87)

Prompt-ready context windows

buildContext assembles the last N messages plus older messages relevant to the current query — the standard fix for context windows that can't hold the whole conversation:

final window = await memory.buildContext(
  'default',
  userInput,
  recentCount: 6,    // always keep the recent tail
  relevantCount: 4,  // plus up to 4 semantically recalled older messages
);

Vector stores

Store	Persistence	Use case
InMemoryVectorStore	session only	tests, ephemeral chat
HiveVectorStore	on-device, survives restarts	production apps

await Hive.initFlutter();
final vectors = await HiveVectorStore.open('agent_memory');

Both support search(topK:, minScore:, filter:), metadata filtering, and deleteWhere. Implement VectorStore to back it with sqlite-vec, ObjectBox, or a server.

Embedding providers

GeminiEmbeddingProvider(apiKey: '...')                 // text-embedding-004
OpenAIEmbeddingProvider(apiKey: '...')                 // text-embedding-3-small
OpenAIEmbeddingProvider(apiKey: '...', baseUrl: '...') // Azure / proxy / local

Implement EmbeddingProvider (one method) to plug in any other model — including on-device embedders.

Direct vector search (RAG building block)

You don't need SemanticMemory to use the primitives:

final store = InMemoryVectorStore();
await store.add(VectorRecord(
  id: 'doc1',
  text: 'Flutter widgets are built with composition.',
  embedding: await embedder.embed('Flutter widgets are built with composition.'),
  metadata: {'source': 'docs'},
));

final hits = await store.search(await embedder.embed('how do widgets work?'),
    topK: 3, minScore: 0.5);

License

MIT — see LICENSE.