Letta

Letta (formerly MemGPT) tackles the problem most agents ignore: they forget everything the moment the conversation ends. I've watched colleagues restart the same context with ChatGPT every Monday because it has no idea what happened last week. Letta gives agents persistent memory that survives sessions, which sounds minor until you need an agent that actually learns your codebase conventions or remembers your project constraints across weeks. The core tech is a memory management system that decides what to keep in active context, what to archive, and what to retrieve when relevant. This isn't just RAG with extra steps. The agent can edit its own memory, tag concepts, and build a knowledge graph over time. I used it to build a documentation assistant that learned our internal API patterns after a few sessions and started suggesting consistent examples without me re-explaining the architecture every time. Letta Code is the newer piece, a local coding agent that uses this memory system. It's slower than Cursor or Aider because it thinks before acting, but that's the point. When I gave it a refactoring task across twelve files, it remembered which patterns I'd rejected in previous sessions and avoided them. That kind of continuity matters more than speed when you're iterating on architecture. The failure mode is predictable: if the agent's memory fills with noise, it starts hallucinating connections that don't exist. I had to manually prune its memory after it decided two unrelated bugs were part of the same root cause because they both involved date parsing. The pruning UI exists but feels like an afterthought. Compared to AutoGPT or similar, Letta is less about autonomous task execution and more about building an agent that improves with use. If you want something to run off and complete a task unsupervised, this isn't it. If you want an agent that gets better at your specific workflows over weeks, it's the only serious option I've found. The freemium model is fair: OSS for self-hosting, paid cloud if you want managed infrastructure. The cloud tier makes sense for teams who don't want to run their own Postgres and vector store.