hermes-brain-memory

Brain Memory is a memory system for AI agents, built around a hierarchical file-system structure. It stores memories as plain Markdown files with YAML frontmatter, and it is described as a memory architecture rather than a vector store. The system is meant to help agents keep persistent memory across sessions and across different models or agent environments.

Its core behavior centers on decay, recall, reinforcement, and consolidation. Memories decay along an exponential Ebbinghaus-style forgetting curve, while deterministic recall combines TF-IDF, decayed strength, and spreading activation from weighted links between related memories. Recalling a memory can strengthen it again through spaced reinforcement, and the product also includes a nine-phase nightly cycle that handles replay, consolidation, pruning, reorganization, REM-style recombination, synaptic homeostasis, knowledge propagation, semantic crystallization, and expertise detection. The page also says that episodic, semantic, and procedural memories each use their own decay rates and consolidation rules.

Brain Memory is designed for cross-agent use. A hosted MCP connector reaches Claude Code, Codex CLI, OpenCode, Copilot CLI, Kilo, Antigravity, and the Claude and ChatGPT apps, and the page says that OpenClaw and Hermes are native memory engines. It is also described as working with any underlying language model. The system can be browsed in any file explorer because the directory tree itself is the semantic structure, and it supports synchronization by pointing a BRAIN_DIR folder at Google Drive, Dropbox, or iCloud, or by using git or encrypted export. No account is required.

The package is available in beta on npm with an install command shown as npm i -g brain-memory@beta. It is licensed under MIT.