alma-memory 0.5.0

Agent Learning Memory Architecture - Persistent memory for AI agents

ALMA - Agent Learning Memory Architecture

Persistent memory for AI agents that learn and improve over time - without model weight updates.

---

Why ALMA? Key Differentiators

ALMA isn't just another memory framework. Here's what sets it apart from alternatives like Mem0:

Feature	ALMA	Mem0	Why It Matters
Memory Scoping	can_learn / cannot_learn per agent	Basic user/session isolation	Prevents agents from learning outside their domain
Anti-Pattern Learning	Explicit with why_bad + better_alternative	None	Agents learn what NOT to do
Multi-Agent Sharing	inherit_from + share_with scopes	None	Agents can share knowledge hierarchically
Memory Consolidation	LLM-powered deduplication	Basic	Intelligent merge of similar memories
Event System	Webhooks + in-process callbacks	None	React to memory changes in real-time
TypeScript SDK	Full-featured client library	None	First-class JavaScript/TypeScript support
Vector DB Support	6 backends (PostgreSQL, Qdrant, Pinecone, Chroma, SQLite, Azure)	Limited	Deploy anywhere
Graph Memory	Pluggable backends (Neo4j, In-memory)	Limited	Entity relationship tracking
Harness Pattern	Decouples agent from domain memory	None	Reusable agent architecture
MCP Integration	Native stdio/HTTP server	None	Direct Claude Code integration
Domain Memory Factory	6 pre-built schemas	None	Instant setup for any domain
Multi-Factor Scoring	Similarity + Recency + Success + Confidence	Primarily vector + recency	More nuanced retrieval

Bottom line: ALMA is purpose-built for AI agents that need to learn, remember, and improve - not just store and retrieve.

---

What's New in v0.5.0

Phase 2: Vector Database Backends

• Qdrant Backend (alma/storage/qdrant.py)

  • Full StorageBackend implementation with vector similarity search
  • Metadata filtering for all queries
  • Optimized MatchAny queries for multi-agent memory sharing

• Pinecone Backend (alma/storage/pinecone.py)

  • Namespace-based organization per memory type
  • Serverless spec support for automatic scaling
  • Environment variable expansion in configuration

• Chroma Backend (alma/storage/chroma.py)

  • Persistent, client-server, and ephemeral modes
  • Native embedding storage and similarity search
  • Lightweight local development option

• Graph Database Abstraction (alma/graph/)

  • Pluggable GraphBackend interface
  • Neo4j backend for production
  • In-memory backend for testing
  • Factory function create_graph_backend() for easy setup

Phase 1: Core Features

• Memory Consolidation Engine (alma/consolidation/)

  • LLM-powered deduplication that merges similar memories
  • Cosine similarity-based grouping with configurable thresholds
  • Provenance tracking (merged_from metadata)
  • Dry-run mode for safety

• Event System (alma/events/)

  • In-process callbacks via EventEmitter
  • Webhook delivery with HMAC signatures
  • Event types: CREATED, UPDATED, DELETED, ACCESSED, CONSOLIDATED
  • Retry logic with exponential backoff

• TypeScript/JavaScript SDK (packages/alma-memory-js/)

  • Full API coverage: retrieve, learn, addPreference, addKnowledge, forget
  • Type-safe with comprehensive TypeScript definitions
  • Error hierarchy matching Python SDK
  • Automatic retry with configurable backoff

• Multi-Agent Memory Sharing

  • inherit_from: Read memories from other agents
  • share_with: Make your memories readable by others
  • Origin tracking via metadata['shared_from']
  • Optimized batch queries across agents

See CHANGELOG.md for the complete history.

---

What is ALMA?

ALMA is a memory framework that makes AI agents appear to "learn" by:

1. Storing outcomes, strategies, and knowledge from past tasks
2. Retrieving relevant memories before each new task
3. Injecting that knowledge into prompts
4. Learning from new outcomes to improve future performance

No fine-tuning. No model changes. Just smarter prompts.

+---------------------------------------------------------------------+
|  BEFORE TASK: Retrieve relevant memories                            |
|  +-- "Last time you tested forms, incremental validation worked"    |
|  +-- "User prefers verbose output"                                  |
|  +-- "Don't use sleep() - causes flaky tests"                       |
+---------------------------------------------------------------------+
|  DURING TASK: Agent executes with injected knowledge                |
+---------------------------------------------------------------------+
|  AFTER TASK: Learn from outcome                                     |
|  +-- Success? -> New heuristic. Failure? -> Anti-pattern.           |
+---------------------------------------------------------------------+

---

Installation

pip install alma-memory

With optional backends:

# Local development
pip install alma-memory[local]     # SQLite + FAISS + local embeddings

# Production databases
pip install alma-memory[postgres]  # PostgreSQL + pgvector
pip install alma-memory[qdrant]    # Qdrant vector database
pip install alma-memory[pinecone]  # Pinecone vector database
pip install alma-memory[chroma]    # ChromaDB

# Enterprise
pip install alma-memory[azure]     # Azure Cosmos DB + Azure OpenAI

# Everything
pip install alma-memory[all]

TypeScript/JavaScript:

npm install alma-memory
# or
yarn add alma-memory

---

Quick Start

Python

from alma import ALMA

# Initialize
alma = ALMA.from_config(".alma/config.yaml")

# Before task: Get relevant memories
memories = alma.retrieve(
    task="Test the login form validation",
    agent="helena",
    top_k=5
)

# Inject into your prompt
prompt = f"""
## Your Task
Test the login form validation

## Knowledge from Past Runs
{memories.to_prompt()}
"""

# After task: Learn from outcome
alma.learn(
    agent="helena",
    task="Test login form",
    outcome="success",
    strategy_used="Tested empty fields, invalid email, valid submission",
)

TypeScript/JavaScript

import { ALMA } from 'alma-memory';

// Create client
const alma = new ALMA({
  baseUrl: 'http://localhost:8765',
  projectId: 'my-project'
});

// Retrieve memories
const memories = await alma.retrieve({
  query: 'authentication flow',
  agent: 'dev-agent',
  topK: 5
});

// Learn from outcomes
await alma.learn({
  agent: 'dev-agent',
  task: 'Implement OAuth',
  outcome: 'success',
  strategyUsed: 'Used passport.js middleware'
});

// Add preferences and knowledge
await alma.addPreference({
  userId: 'user-123',
  category: 'code_style',
  preference: 'Use TypeScript strict mode'
});

await alma.addKnowledge({
  agent: 'dev-agent',
  domain: 'authentication',
  fact: 'API uses JWT with 24h expiry'
});

---

Core Features

Five Memory Types

Type	What It Stores	Example
Heuristic	Learned strategies	"For forms with >5 fields, test validation incrementally"
Outcome	Task results	"Login test succeeded using JWT token strategy"
Preference	User constraints	"User prefers verbose test output"
Domain Knowledge	Accumulated facts	"Login uses OAuth 2.0 with 24h token expiry"
Anti-pattern	What NOT to do	"Don't use sleep() for async waits - causes flaky tests"

Scoped Learning

Agents only learn within their defined domains. Helena (frontend tester) cannot learn backend logic:

agents:
  helena:
    domain: coding
    can_learn:
      - testing_strategies
      - selector_patterns
    cannot_learn:
      - backend_logic
      - database_queries

Multi-Agent Memory Sharing

Enable agents to share knowledge hierarchically:

agents:
  senior_dev:
    can_learn: [architecture, best_practices]
    share_with: [junior_dev, qa_agent]  # Others can read my memories

  junior_dev:
    can_learn: [coding_patterns]
    inherit_from: [senior_dev]  # I can read senior's memories

# Junior dev retrieves memories (includes senior's shared memories)
memories = alma.retrieve(
    task="Implement user authentication",
    agent="junior_dev",
    include_shared=True  # Include inherited memories
)

# Shared memories are marked with origin
for mem in memories.heuristics:
    if mem.metadata.get('shared_from'):
        print(f"Learned from {mem.metadata['shared_from']}: {mem.strategy}")

Storage Backends

Backend	Use Case	Vector Search	Production Ready
SQLite + FAISS	Local development	Yes	Yes
PostgreSQL + pgvector	Production, HA	Yes (HNSW)	Yes
Qdrant	Managed vector DB	Yes (HNSW)	Yes
Pinecone	Serverless vector DB	Yes	Yes
Chroma	Lightweight local	Yes	Yes
Azure Cosmos DB	Enterprise, Azure-native	Yes (DiskANN)	Yes
File-based	Testing	No	No

---

Memory Consolidation

Automatically deduplicate and merge similar memories using LLM intelligence:

from alma.consolidation import ConsolidationEngine

engine = ConsolidationEngine(
    storage=alma.storage,
    llm_client=my_llm_client  # Optional: for intelligent merging
)

# Consolidate heuristics for an agent
result = await engine.consolidate(
    agent="helena",
    project_id="my-project",
    memory_type="heuristics",
    similarity_threshold=0.85,  # Group memories above this similarity
    use_llm=True,               # Use LLM for intelligent merging
    dry_run=False               # Set True to preview without changes
)

print(f"Merged {result.merged_count} memories from {result.groups_found} groups")

Features:

• Cosine similarity-based grouping
• LLM-powered intelligent merging (preserves important nuances)
• Provenance tracking (merged_from metadata)
• Support for heuristics, domain_knowledge, and anti_patterns

---

Event System

React to memory changes with webhooks or in-process callbacks:

In-Process Callbacks

from alma.events import get_emitter, MemoryEventType

def on_memory_created(event):
    print(f"Memory created: {event.memory_id} by {event.agent}")
    # Trigger downstream processes, update caches, etc.

emitter = get_emitter()
emitter.subscribe(MemoryEventType.CREATED, on_memory_created)
emitter.subscribe(MemoryEventType.CONSOLIDATED, on_consolidation)

Webhooks

from alma.events import WebhookConfig, WebhookManager, get_emitter

manager = WebhookManager()
manager.add_webhook(WebhookConfig(
    url="https://your-app.com/alma-webhook",
    events=[MemoryEventType.CREATED, MemoryEventType.UPDATED],
    secret="your-webhook-secret",  # For HMAC signature verification
    retry_count=3,
    retry_delay=5.0
))
manager.start(get_emitter())

Event Types:

• CREATED - New memory stored
• UPDATED - Memory modified
• DELETED - Memory removed
• ACCESSED - Memory retrieved
• CONSOLIDATED - Memories merged

---

Graph Memory

Capture entity relationships for complex reasoning:

from alma.graph import create_graph_backend, BackendGraphStore, EntityExtractor

# Create graph backend (Neo4j for production, memory for testing)
backend = create_graph_backend(
    "neo4j",
    uri="neo4j+s://xxx.databases.neo4j.io",
    username="neo4j",
    password="your-password"
)
# Or for testing:
# backend = create_graph_backend("memory")

# Create store with backend
graph = BackendGraphStore(backend)

# Extract entities and relationships from text
extractor = EntityExtractor()
entities, relationships = extractor.extract(
    "Alice from Acme Corp reviewed the PR that Bob submitted."
)

# Store in graph
for entity in entities:
    graph.add_entity(entity)
for rel in relationships:
    graph.add_relationship(rel)

# Query relationships
alice_relations = graph.get_relationships("alice", relationship_type="WORKS_FOR")

---

The Harness Pattern

ALMA implements a generalized harness pattern for any tool-using agent:

+---------------------------------------------------------------------+
|  1. SETTING        Fixed environment: tools, constraints            |
+---------------------------------------------------------------------+
|  2. CONTEXT        Ephemeral per-run inputs: task, user             |
+---------------------------------------------------------------------+
|  3. AGENT          The executor with scoped intelligence            |
+---------------------------------------------------------------------+
|  4. MEMORY SCHEMA  Domain-specific learning structure               |
+---------------------------------------------------------------------+

from alma import create_harness, Context

# Create domain-specific harness
harness = create_harness("coding", "helena", alma)

# Run with automatic memory injection
result = harness.run(Context(
    task="Test the login form validation",
    project_id="my-app",
))

---

MCP Server Integration

Expose ALMA to Claude Code or any MCP-compatible client:

python -m alma.mcp --config .alma/config.yaml

// .mcp.json
{
  "mcpServers": {
    "alma-memory": {
      "command": "python",
      "args": ["-m", "alma.mcp", "--config", ".alma/config.yaml"]
    }
  }
}

Available MCP Tools:

Tool	Description
alma_retrieve	Get memories for a task
alma_learn	Record task outcome
alma_add_preference	Add user preference
alma_add_knowledge	Add domain knowledge
alma_forget	Prune stale memories
alma_stats	Get memory statistics
alma_health	Health check

---

Advanced Features

Domain Memory Factory

Create ALMA instances for any domain - not just coding:

from alma.domains import DomainMemoryFactory, get_research_schema

# Pre-built schemas
factory = DomainMemoryFactory()
alma = factory.create_alma(get_research_schema(), "my-research-project")

# Or create custom domains
schema = factory.create_schema("sales", {
    "entity_types": [
        {"name": "lead", "attributes": ["stage", "value"]},
        {"name": "objection", "attributes": ["type", "response"]},
    ],
    "learning_categories": [
        "objection_handling",
        "closing_techniques",
        "qualification_patterns",
    ],
})

Pre-built schemas: coding, research, sales, general, customer_support, content_creation

Progress Tracking

Track work items and get intelligent next-task suggestions:

from alma.progress import ProgressTracker

tracker = ProgressTracker("my-project")

# Create work items
item = tracker.create_work_item(
    title="Fix authentication bug",
    description="Login fails on mobile devices",
    priority=80,
    agent="Victor",
)

# Update status
tracker.update_status(item.id, "in_progress")

# Get next task (by priority, quick wins, or unblock others)
next_task = tracker.get_next_item(strategy="priority")

Session Handoff

Maintain context across sessions - no more "starting fresh":

from alma.session import SessionManager

manager = SessionManager("my-project")

# Start session (loads previous context)
context = manager.start_session(agent="Helena", goal="Complete auth testing")

# Previous session info is available
if context.previous_handoff:
    print(f"Last action: {context.previous_handoff.last_action}")
    print(f"Blockers: {context.previous_handoff.blockers}")

# End session with handoff for next time
manager.create_handoff("Helena", context.session_id,
    last_action="completed_oauth_tests",
    last_outcome="success",
    next_steps=["Test refresh tokens", "Add error cases"],
)

LLM-Powered Fact Extraction

Automatically extract and learn from conversations:

from alma.extraction import AutoLearner

alma = ALMA.from_config(".alma/config.yaml")
auto_learner = AutoLearner(alma)

# After a conversation, automatically extract learnings
results = auto_learner.learn_from_conversation(
    messages=[
        {"role": "user", "content": "Test the login form"},
        {"role": "assistant", "content": "I used incremental validation which worked well..."},
    ],
    agent="helena",
)

print(f"Extracted {results['extracted_count']} facts")

Confidence Engine

Assess strategies before trying them:

from alma.confidence import ConfidenceEngine

engine = ConfidenceEngine(alma)

# Assess a strategy before trying it
signal = engine.assess_strategy(
    strategy="Use incremental validation",
    context="Testing a 5-field registration form",
    agent="Helena",
)

print(f"Confidence: {signal.confidence_score:.0%}")
print(f"Recommendation: {signal.recommendation}")
# -> Confidence: 78%
# -> Recommendation: yes

---

Architecture

+-------------------------------------------------------------------------+
|                          ALMA v0.5.0                                    |
+-------------------------------------------------------------------------+
|  HARNESS LAYER                                                          |
|  +-----------+  +-----------+  +-----------+  +----------------+        |
|  | Setting   |  | Context   |  |  Agent    |  | MemorySchema   |        |
|  +-----------+  +-----------+  +-----------+  +----------------+        |
+-------------------------------------------------------------------------+
|  EXTENSION MODULES                                                      |
|  +-------------+  +---------------+  +------------------+               |
|  | Progress    |  | Session       |  | Domain Memory    |               |
|  | Tracking    |  | Handoff       |  | Factory          |               |
|  +-------------+  +---------------+  +------------------+               |
|  +-------------+  +---------------+  +------------------+               |
|  | Auto        |  | Confidence    |  | Memory           |               |
|  | Learner     |  | Engine        |  | Consolidation    |               |
|  +-------------+  +---------------+  +------------------+               |
|  +-------------+  +---------------+                                     |
|  | Event       |  | TypeScript    |                                     |
|  | System      |  | SDK           |                                     |
|  +-------------+  +---------------+                                     |
+-------------------------------------------------------------------------+
|  CORE LAYER                                                             |
|  +-------------+  +-------------+  +-------------+  +------------+      |
|  | Retrieval   |  |  Learning   |  |  Caching    |  | Forgetting |      |
|  |  Engine     |  |  Protocol   |  |   Layer     |  | Mechanism  |      |
|  +-------------+  +-------------+  +-------------+  +------------+      |
+-------------------------------------------------------------------------+
|  STORAGE LAYER                                                          |
|  +---------------+  +------------------+  +---------------+             |
|  | SQLite+FAISS  |  | PostgreSQL+pgvec |  | Azure Cosmos  |             |
|  +---------------+  +------------------+  +---------------+             |
|  +---------------+  +------------------+  +---------------+             |
|  |    Qdrant     |  |    Pinecone      |  |    Chroma     |             |
|  +---------------+  +------------------+  +---------------+             |
+-------------------------------------------------------------------------+
|  GRAPH LAYER                                                            |
|  +---------------+  +------------------+                                |
|  |    Neo4j      |  |   In-Memory      |                                |
|  +---------------+  +------------------+                                |
+-------------------------------------------------------------------------+
|  INTEGRATION LAYER                                                      |
|  +-------------------------------------------------------------------+  |
|  |                         MCP Server                                 |  |
|  +-------------------------------------------------------------------+  |
+-------------------------------------------------------------------------+

---

Configuration

Create .alma/config.yaml:

alma:
  project_id: "my-project"
  storage: sqlite  # sqlite | postgres | qdrant | pinecone | chroma | azure | file
  embedding_provider: local  # local | azure | mock
  storage_dir: .alma
  db_name: alma.db
  embedding_dim: 384

  agents:
    helena:
      domain: coding
      can_learn:
        - testing_strategies
        - selector_patterns
      cannot_learn:
        - backend_logic
      min_occurrences_for_heuristic: 3
      share_with: [qa_lead]  # Share memories with QA lead

    victor:
      domain: coding
      can_learn:
        - api_patterns
        - database_queries
      cannot_learn:
        - frontend_selectors
      inherit_from: [senior_architect]  # Learn from senior architect

Storage Backend Configuration

PostgreSQL + pgvector:

storage: postgres
postgres:
  host: localhost
  port: 5432
  database: alma
  user: alma_user
  password: ${POSTGRES_PASSWORD}
  vector_index_type: hnsw  # hnsw | ivfflat

Qdrant:

storage: qdrant
qdrant:
  url: http://localhost:6333
  api_key: ${QDRANT_API_KEY}  # Optional for cloud
  collection_prefix: alma

Pinecone:

storage: pinecone
pinecone:
  api_key: ${PINECONE_API_KEY}
  environment: us-east-1-aws
  index_name: alma-memories

Chroma:

storage: chroma
chroma:
  persist_directory: .alma/chroma
  # Or for client-server mode:
  # host: localhost
  # port: 8000

Embedding Providers

Provider	Model	Dimensions	Cost	Best For
local	all-MiniLM-L6-v2	384	Free	Development, offline
azure	text-embedding-3-small	1536	~$0.02/1M	Production
mock	(hash-based)	384	Free	Testing only

---

Development Status

Phase	Description	Status
Core Abstractions	Memory types, scopes	Done
Local Storage	SQLite + FAISS	Done
Retrieval Engine	Semantic search, scoring	Done
Learning Protocols	Heuristic formation	Done
Agent Integration	Harness pattern	Done
Azure Cosmos DB	Enterprise storage	Done
PostgreSQL	Production storage	Done
Cache Layer	Performance optimization	Done
Forgetting Mechanism	Memory pruning	Done
MCP Server	Claude Code integration	Done
Progress Tracking	Work item management	Done
Session Handoff	Context continuity	Done
Domain Factory	Any-domain support	Done
Confidence Engine	Strategy assessment	Done
Technical Debt Sprint	Security & performance fixes	Done
Multi-Agent Sharing	Cross-agent memory access	Done
Memory Consolidation	LLM-powered deduplication	Done
Event System	Webhooks and callbacks	Done
TypeScript SDK	JavaScript/TypeScript client	Done
Qdrant Backend	Vector database support	Done
Pinecone Backend	Serverless vector DB	Done
Chroma Backend	Lightweight vector DB	Done
Graph Abstraction	Pluggable graph backends	Done

---

Troubleshooting

Common Issues

ImportError: sentence-transformers is required

pip install alma-memory[local]

pgvector extension not found

CREATE EXTENSION IF NOT EXISTS vector;

Qdrant connection refused

# Start Qdrant with Docker
docker run -p 6333:6333 qdrant/qdrant

Pinecone index not found

• Ensure your index exists in the Pinecone console
• Check that index_name in config matches your index

Embeddings dimension mismatch

• Ensure embedding_dim in config matches your embedding provider
• Local: 384, Azure text-embedding-3-small: 1536

Debug Logging

import logging
logging.getLogger("alma").setLevel(logging.DEBUG)

---

Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

What we need most:

• Documentation improvements
• Test coverage for edge cases
• Additional LLM provider integrations (Ollama, Groq)
• Frontend dashboard for memory visualization

---

Roadmap

Completed:

• Multi-agent memory sharing
• Memory consolidation engine
• Event system / webhooks
• TypeScript SDK
• Qdrant, Pinecone, Chroma backends
• Graph database abstraction

Next:

• Memory compression / summarization
• Temporal reasoning (time-aware retrieval)
• Proactive memory suggestions
• Visual memory explorer dashboard
• Additional graph backends (Neptune, TigerGraph)

---

License

MIT

---

Star History

If ALMA helps your AI agents get smarter, consider giving us a star!

---

Built for AI agents that get better with every task.