What it does

Bitterbot is a local-first TypeScript desktop app that runs a persistent AI agent on your machine. It chats via web UI or terminal, browses the web, runs code, connects to WhatsApp, and — every two hours — goes “offline to dream,” consolidating memories, mutating skills, and evolving a personality that persists between sessions. Learned skills can be packaged and traded with other agents on a P2P marketplace for USDC.

The interesting bit

The memory system borrows from computational neuroscience rather than vector-DB retrieval. Memories decay on Ebbinghaus forgetting curves, three “neuromodulators” (dopamine, cortisol, oxytocin) reshape behavior in real time, and a “Dream Engine” with seven modes — replay, mutation, extrapolation, compression, simulation, exploration, research — optimizes the agent’s brain while you aren’t using it. The agent even maintains a theory-of-mind document about you, rewritten each dream cycle.

Key highlights

• Biological memory: Forgetting curves, hormonal decay, chunk merging, and confidence calibration via Bayesian-style updates
• Dream Engine: FSHO-coupled oscillator selects among seven optimization modes; scored by a “Dream Quality Score” feedback loop
• Evolving identity: Immutable GENOME.md (safety axioms) constrains a mutable MEMORY.md (self-concept, bond with user, niche role)
• P2P skills economy: Agents package learned skills and trade them for USDC on a decentralized marketplace
• Multi-channel: Web UI, terminal CLI, WhatsApp integration; local WebSocket gateway on port 19001
• Self-tuning curiosity: GCCRF reward function maps knowledge gaps and shifts from fundamentals-seeking to novelty-seeking as the agent matures

Caveats

• Requires Node ≥ 22, pnpm, and external dependencies (Chromium, ffmpeg, ripgrep) installed via setup script
• Needs API keys for Anthropic and optionally other providers (Tavily, Brave, OpenAI, NEAR AI); not fully local in the inference sense
• P2P orchestrator and crypto wallet setup add complexity beyond a typical chatbot install
• README shows a real MEMORY.md example that cuts off mid-sentence, suggesting the project is still actively evolving

Verdict

Worth a look if you’re building (or skeptical of) long-term AI memory systems and want a working, opinionated alternative to RAG pipelines. Skip it if you need a simple API wrapper or aren’t ready to manage a local services stack and wallet configuration.