memlocal_dart

local AI memory for Flutter apps. give your LLM a brain that remembers across sessions — facts, preferences, relationships, and context — all stored on-device with CozoDB.

website · GitHub · example app · API docs

what is memlocal?

memlocal is a local-first AI memory layer modelled on human cognitive architecture. it gives any Flutter + LLM app the ability to:

remember facts, preferences, experiences, and relationships across sessions
search memories with vector similarity, full-text BM25, graph traversal, or hybrid
inject context automatically into LLM prompts
extract memories from conversations using LLM-driven inference
build a knowledge graph that connects related memories

all data stays on-device in an embedded CozoDB database — no cloud sync required.

memory architecture

memlocal mirrors the three-tier human memory model:

tier	types	TTL
sensory	sensory buffer	~5 seconds
short-term	working memory, attention context, conversation buffer	~5 minutes
long-term	episodic, semantic, factual, procedural, social, spatial, prospective, affective	permanent

12 memory subtypes are automatically classified during extraction so you can filter, weight, and reason over different categories of knowledge.

features

LLM-driven extraction — automatically parse facts, preferences, events, and relationships from conversations
multi-modal search — semantic (HNSW vector), full-text (BM25), graph traversal, and hybrid search with re-ranking
knowledge graph — 10 relationship types (relatesTo, contradicts, supersedes, causedBy, partOf, prefersOver, follows, instanceOf, belongsTo, similarTo)
user profiles — automatically maintained from extracted memories (static facts + dynamic context)
prospective memory — reminders triggered by topic, time, user presence, or semantic match
context builder — assembles a ready-to-inject context block from relevant memories, profile, and attention items
tool calling — 9 native tools formatted for Claude, OpenAI, and Gemini APIs
deduplication — content-hash dedup with LLM-driven update/merge for conflicting facts
on-device — everything stored locally in CozoDB with SQLite persistence
multi-provider — works with Anthropic, OpenAI, and Google embedding + LLM APIs

supported providers

LLMs (for extraction)

provider	models
Anthropic	Claude Sonnet 4.6, Claude Haiku 4.5
OpenAI	GPT-5.2, GPT-5 Mini
Google	Gemini 3.1 Pro, Gemini 3 Flash
custom	any OpenAI-compatible endpoint

embeddings (for vector search)

provider	model	dimensions
OpenAI	`text-embedding-3-small`	1536
Google	`gemini-embedding-001`	768
custom	any compatible endpoint	configurable

note: Anthropic does not offer an embedding API. when using Anthropic LLMs, memlocal defaults to OpenAI embeddings — provide an OpenAI API key via embeddingApiKey.

getting started

1. install

dependencies:
  memlocal_dart: ^1.0.0

flutter pub get

2. initialize

import 'package:memlocal_dart/memlocal_dart.dart';

final mem = await memlocal.init(
  memlocalConfig.withDefaults(
    provider: LlmProvider.claudeSonnet4_6,
    apiKey: 'sk-ant-...',
    embeddingApiKey: 'sk-openai-...', // Required for Anthropic
  ),
);

3. add memories

// automatic extraction — LLM parses facts from text
final result = await mem.add(
  'My name is Sirsho, I live in Bengaluru, and I love hiking.',
  userId: 'sirsho_01',
);
// result.items → [FactualMemory("name: Sirsho"), SpatialMemory("lives in Bengaluru"), ...]

// direct storage — no LLM inference
await mem.add(
  'Prefers dark mode',
  userId: 'sirsho_01',
  type: MemoryType.factual,
  infer: false,
);

4. search

// semantic search (vector similarity)
final results = await mem.search(
  'Where does the user live?',
  mode: SearchMode.semantic,
  userId: 'sirsho_01',
);

// full-text search (BM25)
final textResults = await mem.search(
  'Bengaluru',
  mode: SearchMode.text,
);

// hybrid (semantic + BM25, merged & re-ranked)
final hybrid = await mem.search(
  'outdoor hobbies near Bengaluru',
  mode: SearchMode.hybrid,
  userId: 'sirsho_01',
  k: 5,
);

5. context injection

the fastest way to make your LLM context-aware:

// build a context block for the current query
final context = await mem.getContext(
  'What should we do this weekend?',
  userId: 'sirsho_01',
);

// inject into your system prompt
final systemPrompt = '''You are a helpful assistant.

${context.contextBlock}''';
// contextBlock contains relevant memories grouped by type, user profile, and attention items

6. conversation buffer

track the conversation for extraction:

await mem.addMessage(Message.user('I just ran a 10k in Forest Park!'));
await mem.addMessage(Message.assistant('That sounds great! How was the trail?'));

// retrieve recent messages
final messages = await mem.getConversation(limit: 20);

7. knowledge graph

build relationships between memories:

await mem.addEdge(MemoryEdge(
  fromId: runningMemoryId,
  toId: BengaluruMemoryId,
  relation: MemoryRelation.relatesTo,
  weight: 0.9,
));

8. tool calling

expose memory tools to your LLM for agentic flows:

// get tool definitions in provider-native format
final tools = mem.toolDefinitions;
// pass these to your LLM API call as the `tools` parameter

// when the LLM returns tool calls, execute them:
final results = await mem.executeTools(response.toolCalls);

available tools: add_memory, search_memory, get_memories, delete_memory, get_user_profile, add_relationship, get_relationships, add_reminder, get_context

configuration

full configuration

final config = memlocalConfig(
  llm: LlmConfig(
    provider: LlmProvider.claudeSonnet4_6,
    apiKey: 'sk-ant-...',
    temperature: 0.1,          // low temp for reliable extraction
    maxTokens: 1024,
    modelOverride: null,       // use provider default
    baseUrlOverride: null,     // use provider default
  ),
  embedding: EmbeddingConfig(
    provider: EmbeddingProvider.openAi,
    apiKey: 'sk-openai-...',
    modelOverride: null,       // text-embedding-3-small
    dimensionsOverride: null,  // 1536
  ),
  storage: StorageConfig(
    inMemory: false,           // SQLite persistence
    dbPath: null,              // auto via path_provider
    hnswM: 16,                 // HNSW graph connectivity
    hnswEfConstruction: 100,   // HNSW construction quality
  ),
  conversationBufferSize: 20,  // keep last 20 messages
  sensoryBufferCapacity: 100,
  sensoryTtl: Duration(seconds: 5),
  customExtractionPrompt: null,  // override the extraction prompt
  customUpdatePrompt: null,      // override the update/merge prompt
);

final mem = await memlocal.init(config);

quick setup (recommended)

// auto-configures embedding provider based on LLM choice
final mem = await memlocal.init(
  memlocalConfig.withDefaults(
    provider: LlmProvider.claudeSonnet4_6,
    apiKey: 'sk-ant-...',
    embeddingApiKey: 'sk-openai-...',
    storage: StorageConfig(inMemory: false),
  ),
);

how extraction works

when you call mem.add(content, infer: true):

LLM extraction — the content is sent to the configured LLM with a structured prompt that asks it to identify facts, preferences, events, and relationships
candidate parsing — the LLM's JSON response is parsed into candidate MemoryItems with appropriate MemoryType classification
deduplication — each candidate is checked against existing memories by content hash. duplicates are skipped; conflicts are sent back to the LLM for reconciliation
embedding — new/updated items are batch-embedded via the embedding API
storage — items are upserted into CozoDB with their vectors, and HNSW/FTS/LSH indices are automatically maintained

the entire pipeline runs locally except for the LLM and embedding API calls.

advanced usage

access the memory manager

for fine-grained control over individual memory subsystems:

final manager = mem.manager;

// direct access to subsystems
await manager.semanticMemory.record(item, embedding);
await manager.episodicMemory.record(item, embedding);
await manager.factualMemory.record(item, embedding);
await manager.socialMemory.record(item, embedding);

// working memory control
manager.workingMemory.setRelevant(items);
manager.workingMemory.focus(importantItem);
final block = manager.workingMemory.toContextBlock();

// storage layer
final store = manager.store;
await store.putEdge(edge);
final ranks = await store.pageRank();
final communities = await store.communityDetection();

user profiles

// profiles are auto-maintained from extracted memories
final profile = await mem.getProfile('sirsho_01');
print(profile?.staticFacts);  // {name: Sirsho, location: Bengaluru, ...}
print(profile?.dynamicContext); // {recent_topic: trail running, ...}
print(profile?.toSummary());  // Formatted text for LLM context

// manual update
await mem.updateProfile(profile!.copyWith(
  staticFacts: {...profile.staticFacts, 'occupation': 'Engineer'},
));

prospective memory (reminders)

// create a reminder triggered by topic mention
await mem.manager.store.putProspective(ProspectiveItem(
  id: 'reminder_1',
  content: 'Ask about the Bengaluru marathon',
  triggerType: TriggerType.topicMention,
  triggerCondition: 'marathon OR running race',
  userId: 'sirsho_01',
));

// check pending reminders
final pending = await mem.manager.store.getPendingProspective(userId: 'sirsho_01');

API reference

`memlocal` (main facade)

method	description
`memlocal.init(config)`	initialize the memory system
`add(content, ...)`	add memory (with optional LLM inference)
`addMessage(message)`	add to conversation buffer
`addEdge(edge)`	add a knowledge graph edge
`search(query, ...)`	search memories (semantic / text / graph / hybrid)
`getContext(query, ...)`	build assembled context for LLM injection
`getProfile(userId)`	get auto-maintained user profile
`updateProfile(profile)`	update user profile
`getConversation(...)`	get recent conversation messages
`clearConversation(...)`	clear conversation buffer
`toolDefinitions`	get LLM tool definitions (provider-native format)
`executeTools(toolCalls)`	execute LLM tool calls
`memoryCount`	total stored memories
`dispose()`	clean up resources

memory types

type	category	description
`episodic`	long-term	personal experiences and events
`semantic`	long-term	general knowledge and concepts
`factual`	long-term	personal facts, preferences, profile data
`procedural`	long-term	how-to knowledge, workflows
`social`	long-term	relationships and social context
`spatial`	long-term	locations, places, spatial context
`prospective`	long-term	future intentions, reminders, goals
`affective`	long-term	emotions, sentiments, moods
`workingMemory`	short-term	active context for current query
`attentionContext`	short-term	explicitly focused items
`conversationBuffer`	short-term	recent conversation messages
`sensoryBuffer`	sensory	raw incoming signals (~5s TTL)

search modes

mode	strategy
`SearchMode.semantic`	HNSW vector similarity search
`SearchMode.text`	BM25 full-text search
`SearchMode.graph`	graph traversal from nearest match
`SearchMode.hybrid`	all modes merged and re-ranked

LLM tools

tool	description
`add_memory`	store a fact, preference, or experience
`search_memory`	search memories by query
`get_memories`	list memories with optional filters
`delete_memory`	delete a memory by ID
`get_user_profile`	get user profile summary
`add_relationship`	create a knowledge graph edge
`get_relationships`	get edges connected to a memory
`add_reminder`	create a prospective memory / reminder
`get_context`	get assembled memory context

example app

the example/ directory contains a full-featured chat app demonstrating:

chat with memory-augmented LLM responses
inline memory event timeline (extraction, search, context building)
memory browser with semantic/text/hybrid search
settings page with provider configuration
.env-based configuration

cd example
cp .env.example .env
# Edit .env with your API keys
flutter run

platform support

platform	status
iOS	✅
Android	✅
macOS	✅
Linux	✅
Windows	✅

CozoDB's Rust core is compiled via flutter_rust_bridge through the cozo_dart package.

requirements

Flutter >=3.3.0
Dart SDK ^3.9.2
an API key from Anthropic, OpenAI, or Google (for LLM extraction + embeddings)

license

MIT — see LICENSE for details.

memlocal_dart

what is memlocal?

memory architecture

features

supported providers

LLMs (for extraction)

embeddings (for vector search)

getting started

1. install

2. initialize

3. add memories

4. search

5. context injection

6. conversation buffer

7. knowledge graph

8. tool calling

configuration

full configuration

quick setup (recommended)

how extraction works

advanced usage

access the memory manager

user profiles

prospective memory (reminders)

API reference

`memlocal` (main facade)

memory types

search modes

LLM tools

example app

platform support

requirements

license

links

Libraries

memlocal_dart package

memlocal_dart

what is memlocal?

memory architecture

features

supported providers

LLMs (for extraction)

embeddings (for vector search)

getting started

1. install

2. initialize

3. add memories

4. search

5. context injection

6. conversation buffer

7. knowledge graph

8. tool calling

configuration

full configuration

quick setup (recommended)

how extraction works

advanced usage

access the memory manager

user profiles

prospective memory (reminders)

API reference

memlocal (main facade)

memory types

search modes

LLM tools

example app

platform support

requirements

license

links

Libraries

memlocal_dart package

`memlocal` (main facade)