You don't need a new framework. You need better memory.
Letta (formerly MemGPT) pioneered the idea of letting LLMs manage their own memory. It's a full agent runtime with memory built in. Hebbrix takes a different approach: a standalone memory API that works with whatever you already have.
What you're actually running
The infrastructure behind each approach.
6 layers to manage, scale, and monitor
1 API key. Nothing to host.
Everything below the API layer, the search pipeline, the vector store, the knowledge graph, the embedding models, is managed for you. You focus on your agent. Hebbrix focuses on memory.
Three fundamental differences
Who decides what to remember
In Letta, the LLM itself manages memory through tool calls. It decides what to store in "core memory" (a small editable block in the system prompt) and what to archive. This is elegant in theory, but it means memory quality depends entirely on the model you use. GPT-4 handles it well. Smaller models forget things or store noise.
Hebbrix uses an engineered extraction pipeline. When you send text with infer:true, it extracts clean atomic facts through a 3-tier process: embedding classification, LLM fact extraction, and content deduplication. The quality doesn't depend on which model your agent runs.
How search actually works
Letta's archival memory uses vector similarity search only. Good for semantic meaning, but it misses things. Try searching for "Sarah from TechCorp" and vector search might return results about "professional contacts" or "corporate partnerships" instead of the exact person you mean.
Corporate partnership discussions...
Q2 planning meeting notes...
Relevant, but not the right Sarah
Sarah Chen, VP Engineering at TechCorp
Mentioned Q2 deadline: March 30 for API v2
Exact match via keyword + graph + semantic
Integration cost
Using Letta's memory means running your agents inside Letta's runtime. It has its own agent loops, its own message format, and its own REST API. If you already have agents in LangChain, CrewAI, or your own code, you'd need to rewrite or wrap them.
Hebbrix's chat endpoint is OpenAI compatible. In many cases, adding memory means changing a base URL. Your agent code, your prompts, your orchestration layer, all stay exactly as they are.
When Letta is the right choice
Letta pioneered important ideas. Here's when it genuinely fits better.
When Hebbrix fits better
If any of these describe your situation, Hebbrix is probably the faster path.
Ready to try a different approach?
The free tier gives you 1,000 memories, full hybrid search, and automatic knowledge graphs. Enough to run a real proof of concept.