AI agents often suffer from a lack of persistent memory, forgetting context between sessions and tools. Hippo addresses this by providing a biologically-inspired, zero-dependency memory system that implements decay, retrieval strengthening, and consolidation. It acts as a shared memory layer, enabling AI agents to retain knowledge across different tools and sessions, benefiting multi-tool developers and teams by preventing repeated mistakes and organizing information effectively.

How It Works

Hippo models memory on the human hippocampus, utilizing a three-tiered system: a volatile buffer for current session data, an episodic store for timestamped memories, and a semantic store for consolidated patterns. New information enters the buffer, then gets encoded into episodic memory with assigned tags, strength, and a default half-life. During a "sleep" consolidation phase, repeated episodes are compressed into stable semantic patterns. Decay is default; memories fade unless retrieved, and retrieval strengthens them, mimicking biological memory lifecycle principles.

Quick Start & Requirements

• Primary install: npm install -g hippo-memory
• Initialize: hippo init
• Prerequisites: Node.js 22.5+. Optional embeddings via @xenova/transformers.
• Zero runtime dependencies.

Highlighted Details

• Biologically-inspired memory lifecycle: decay by default, retrieval strengthening, consolidation via "sleep," and reward-proportional decay based on outcome feedback.
• Hybrid search combining BM25 keywords with optional embedding similarity for improved recall.
• Active invalidation for migrating patterns and manual invalidation, alongside confidence tiers (verified, observed, inferred).
• Cross-tool import from ChatGPT, Claude, Cursor, and markdown files, with zero-config agent integration and automatic hook installation.
• Auto-learning from Git commits (hippo learn --git) and failure monitoring (hippo watch) for continuous improvement.
• Features like session handoffs, working memory, and explainable recall enhance agent continuity and transparency.

Maintenance & Community

Issues and PRs are welcomed. The README lists several areas for potential contribution, such as improving LongMemEval scores, developing consolidation heuristics, and building a web UI. No specific community links (e.g., Discord, Slack) or prominent maintainer details are provided.

Licensing & Compatibility

MIT License. This license permits commercial use and integration into closed-source projects without significant restrictions.

Limitations & Caveats

While Hippo achieves a 74.0% Recall@5 on the LongMemEval benchmark using only BM25 (zero dependencies), reaching scores comparable to embedding-based systems requires installing the optional @xenova/transformers dependency. Some areas for future development include enhancing consolidation heuristics and creating a web UI.