Hippo

Hippo is built on a specific critique of most “AI memory” tools: saving everything and bolting on semantic search just means an agent’s memory grows unbounded and noisy, which is why agents keep re-discovering the same known bug repeatedly. Hippo instead models memory on the hippocampus — memories decay by default unless reinforced through use, and every memory carries provenance about where it came from.

Installing it is a single global npm command (npm install -g hippo-memory && hippo init --scan ~) that gives every git repository on the machine a memory layer, working across Claude Code, Codex, Cursor, OpenCode, and any MCP-compatible client. It can also import existing context from ChatGPT, Claude’s CLAUDE.md files, Cursor’s .cursorrules, Slack, or plain markdown, rather than starting from a blank memory.

Storage is SQLite with markdown mirrors, making memory git-trackable and human-readable rather than opaque, and the project has zero runtime dependencies beyond Node.js 22.5+, with optional embeddings via a local transformers model by default or an opt-in API embedder (OpenAI, Voyage, or Cohere) if you want it.

What You Get

• A decay-by-default, strengthen-through-use memory model instead of an ever-growing memory store
• One-command setup (hippo init --scan ~) that adds memory to every git repo on your machine
• Compatibility with Claude Code, Codex, Cursor, OpenCode, Pi, and any MCP client
• Import support from ChatGPT, CLAUDE.md, .cursorrules, Slack, and markdown, so existing context isn’t lost

Common Use Cases

• Preventing an AI coding agent from repeatedly re-discovering the same known bug or deploy issue across sessions
• Migrating existing context from CLAUDE.md, .cursorrules, or ChatGPT history into a structured, decay-aware memory store
• Giving multiple CLI agents (Claude Code, Cursor, Codex) shared memory about the same codebase via a common MCP-accessible store
• Keeping agent memory git-trackable and reviewable as markdown rather than opaque vector database entries

Under The Hood

Architecture Hippo uses SQLite as the memory backbone with markdown file mirrors, so memory state is both queryable structurally and readable/diffable as plain text in git — a deliberate choice to make agent memory auditable rather than a black box. The decay model is core to the design: rather than treating every stored memory as equally permanent, memories lose relevance over time unless reinforced by repeated access, directly modeling how the hippocampus consolidates and forgets.

Tech Stack TypeScript/Node.js (22.5+ required), distributed as the hippo-memory npm package, with zero runtime dependencies for the core system. Embeddings default to a local @xenova/transformers model (no external API needed) with an opt-in path to OpenAI, Voyage, or Cohere embedding APIs for teams that want cloud-based embeddings instead.

Code Quality Very active, consistently maintained commit history and a clear, opinionated design philosophy documented directly in the README (explaining specifically why naive “save everything” memory systems fail) suggest a project built around a considered thesis rather than a generic memory-store clone.

What Makes It Unique Most AI agent memory tools are effectively semantic search over an append-only log; Hippo’s decay-by-default, reinforcement-through-use model plus explicit provenance tracking is a structurally different approach, aimed specifically at preventing the accumulation of stale, unused, or contradictory memories over time.

Self-Hosting

Licensing Model MIT licensed — fully open source with no license key.

Self-Hosting Restrictions Not applicable; Hippo runs entirely locally via a CLI and SQLite, with no hosted service involved unless you opt into an external embedding API.

License Key Required No.