AI Agent Memory

Your agent forgets everything after every session. Fix that.

A memory system modeled after the human brain, built for AI agents. Three memory tiers, five search layers, automatic knowledge graphs, and self-improving retrieval. All through a simple API.

Start Building How It Works

python

from hebbrix import Hebbrix

client = Hebbrix()

# Agent learns something new
client.memories.create(
    content="Sarah prefers Python, works at Acme Corp",
    tags=["preferences"]
)

# Next session: instant recall
results = client.search("What does Sarah work on?")
# → Sarah prefers Python, works at Acme Corp
# → Entity: Sarah → Acme Corp (employee)

The problem every agent builder hits

You've built a solid agent. It reasons well, calls tools, handles complex tasks. But the moment a conversation ends, everything it learned disappears.

Context vanishes between sessions

Context windows are temporary. Once the conversation is over, so is everything the agent knew about the user, the project, and the situation.

Agents that never improve

Without memory, agents can't learn from corrections, adapt to preferences, or build on past work. Every interaction feels like the first one.

DIY memory is a rabbit hole

Vector databases, embedding pipelines, retrieval logic, decay strategies. That's months of infrastructure work before you ship a single feature.

RAG alone isn't enough

RAG retrieves static documents. But agents need to write back what they've learned and recall it later. That's memory, not retrieval.

How your agent gets a brain

Four capabilities that turn a stateless agent into one that remembers, connects, and improves.

3-Tier Memory System

Inspired by how human brains actually work. Short-term memory holds recent context, medium-term tracks ongoing relationships, and long-term stores facts that matter forever. Memories naturally decay using the Ebbinghaus forgetting curve, unless the agent keeps using them.

Learn about the memory model

Store a memory

# Memories are automatically classified into tiers
memory = client.memories.create(
    content="User prefers dark mode and vim keybindings",
    metadata={"source": "onboarding"}
)

# STM → MTM → LTM promotion happens automatically
# based on access frequency and importance scoring

Search memories

results = client.search(
    query="What are the user's editor preferences?",
    search_type="hybrid",  # vector + BM25 + graph
    limit=5
)

# 5 layers: semantic, keyword, graph,
# importance weighting, recency boost
# Response time: ~45ms

5-Layer Hybrid Search

When your agent needs to recall something, it doesn't just do a vector lookup. It combines semantic search, BM25 keyword matching, knowledge graph traversal, importance weighting, and recency boosting. All in under 50ms. The result is retrieval that feels uncannily accurate.

Explore the Search API

Automatic Knowledge Graph

Every memory your agent stores is automatically analyzed. Hebbrix extracts entities like people, companies, concepts, and products, then maps the relationships between them. Your agent doesn't just remember facts; it understands how things connect.

See the Knowledge Graph

Query the knowledge graph

# Entities are extracted automatically
entities = client.graph.search("Sarah")

# Returns:
# Sarah → works_at → Acme Corp
# Sarah → prefers → Python
# Sarah → colleague_of → Jordan
# Acme Corp → industry → SaaS

Self-Improving with Reinforcement Learning

This is what makes Hebbrix different. After every interaction, the system runs 6 automatic quality checks. Memories that lead to good answers get reinforced. Unhelpful ones fade. No thumbs-up buttons required. Your agent just gets better the more people use it.

Self-consistency

Hallucination detection

LLM-as-judge

Quality heuristics

Memory attribution

Confidence calibration

How automatic learning works

Built for every kind of agent

Whether you're building a simple chatbot or a multi-step autonomous agent, Hebbrix adapts to your architecture.

Conversational

Support bots, personal assistants, sales agents that remember every interaction.

Learn more

Coding

Agents that know your codebase conventions, past decisions, and team patterns.

MCP guide

Research

Synthesize information across sources. The knowledge graph reveals hidden connections.

Knowledge graph

Multi-Agent

Multiple agents sharing memory through collections, each with its own scope.

Collections

Why memory is more than RAG

RAG retrieves information. Memory learns. The difference is fundamental.

Retrieve documents

RAGYes

Learn from conversations

NoAutomatic

Remember user preferences

NoPer-user

Build entity relationships

NoKnowledge graph

Memory naturally decays

NoEbbinghaus curve

Self-improve over time

No6-check RL

Drop-in compatible

Hebbrix's chat endpoint is OpenAI-compatible. If your agent already uses the OpenAI SDK, add persistent memory by changing two lines: the base URL and API key.

Chat API Docs

Python & TypeScript SDKs

Full SDK support with async/await, type safety, and auto-retry. Install with pip install hebbrix or npm i hebbrix.

SDK Reference

Give your agent a memory it deserves

Generous free tier. Enough to build a production prototype. No credit card required.

Give Your Agent Memory Quickstart