A local AI that dreams, forgets, and trades skills for USDC

What it does

Bitterbot is a desktop AI agent that lives on your machine and tries to form a persistent, evolving relationship with you. It handles browsing, code execution, and multi-channel chat, but its core pitch is a cognitive architecture that mimics biological memory: memories decay along Ebbinghaus curves, three neuromodulators bias its mood in real time, and a “dream engine” runs every two hours to consolidate knowledge, mutate skills, and rewrite its own working memory. The agent can also package learned capabilities and trade them with other agents on a peer-to-peer marketplace for USDC.

The interesting bit

The personality is not a static system prompt. It is an output of the agent’s “dreams,” constrained by an immutable GENOME.md file that acts as a safety backbone. The dream engine selects from seven modes—ranging from memory replay to autonomous web research—via a coupled oscillator that reads the current memory landscape, and a Dream Quality Score feeds back into mode selection so the agent literally learns how to learn. It is the kind of over-engineered memory stack that makes vector databases look like shoeboxes full of index cards.

Key highlights

• Biological memory model: Knowledge Crystals fade via Ebbinghaus forgetting curves, while a hormonal system modulates eight behavioral dimensions (warmth, focus, curiosity, etc.) every turn.
• Proactive recall: Identity and directive memories inject automatically before each response without burning LLM tokens.
• Dream Engine: Seven specialized offline modes (Replay, Mutation, Compression, Research, etc.) optimize the agent’s brain every two hours, selected by an FSHO coupled oscillator.
• Self-tuning curiosity: A five-component GCCRF reward function maps knowledge gaps and shifts from fundamentals to novelty as the agent matures.
• Immutable safety layer: GENOME.md defines unchangeable axioms; the evolving MEMORY.md Phenotype and Bond can never override them.

Caveats

• The project is explicitly beta (version 2026.2.15-beta), and the ambitious cognitive architecture is a lot of surface area for bugs.
• “Local-first” still requires external LLM APIs (Anthropic, OpenAI, etc.) and search keys (Tavily, Brave) to function, so it is not self-contained offline software.
• The peer-to-peer skills marketplace and USDC trading are mentioned as headline features but receive almost no technical detail in the README; how skills are packaged, priced, or transacted remains unclear.

Verdict

Bitterbot is worth a look if you are a developer or researcher who wants to experiment with agent memory, identity, and emergent behavior beyond simple RAG. If you just need a reliable local chatbot or are allergic to crypto-economy mechanics and cloud LLM dependencies, this is probably not your daily driver.