Skip to content

Hippo Memory

★ New
assess
AI / ML open-source Unknown open-source

At a Glance

Zero-dependency TypeScript CLI and npm package implementing biologically-inspired AI agent memory with exponential decay, retrieval strengthening, episodic-to-semantic consolidation, and SQLite-backed hybrid search.

Type
open-source
Pricing
open-source
License
Unknown
Adoption fit
small
Top alternatives

What It Does

Hippo Memory is a TypeScript CLI tool and npm package that applies biologically-inspired memory mechanics to AI agent session persistence. The core idea is drawn from hippocampal memory theory: memories have a half-life and naturally decay unless retrieved, retrieval strengthens them, and frequently repeated episodic memories consolidate into compressed semantic patterns over time. The project targets developers building with AI coding agents (Claude Code, Cursor, Codex) who want persistent cross-session context without depending on a managed cloud memory service.

The system uses SQLite as its storage backbone with optional @xenova/transformers for embedding-based semantic search, falling back to BM25 keyword search when embeddings are not available. At installation, it auto-detects and patches CLAUDE.md, AGENTS.md, and .cursorrules to inject memory recall and save instructions. It also parses git history to detect migration commits that may render stored memories stale.

Key Features

  • Three-layer memory architecture: Buffer (working memory, no decay), Episodic Store (timestamped entries with 7-day half-life by default), and Semantic Store (consolidated patterns extracted from repeated episodes)
  • Exponential decay with retrieval strengthening: Each memory recall extends half-life by 2 days; error-tagged memories get 2x half-life for sticky learning; unused memories marked stale after 30+ days
  • Episodic-to-semantic consolidation: Frequently accessed episodic memories are compressed into semantic patterns during consolidation passes, reducing storage and improving retrieval signal
  • Hybrid BM25 + embedding search: SQLite FTS5 for keyword search; optional @xenova/transformers for cosine similarity; graceful fallback to BM25 only when embeddings not installed
  • Token-budget-aware context generation: hippo recall accepts a --budget flag (in tokens) and returns memories sized to fit within the specified context window budget
  • Git-based active invalidation: Parses git commit history to detect migration commits and marks related memories as potentially stale
  • Cross-tool auto-patching: Detects and patches CLAUDE.md, AGENTS.md (Codex/OpenCode), and .cursorrules (Cursor) with memory integration instructions
  • Multi-agent memory sharing: Supports peering, sharing, and transfer scoring — universal lessons can travel between projects while project-specific config stays local
  • Zero runtime dependencies (core): Core package ships without mandatory npm dependencies; Node.js 22.5+ required

Use Cases

  • Solo developer coding agent memory: A developer using Claude Code or Cursor who wants errors and decisions remembered across sessions without connecting to a cloud service
  • Project migration tracking: Agents working on refactors where memories about old module locations or deprecated APIs should auto-expire when migration commits are detected
  • Low-budget AI agent setups: Environments where Mem0, Zep, or Weaviate Engram are too expensive or complex, and a local SQLite-backed store suffices
  • Teaching tool: The biological metaphor and decay mechanics make Hippo a useful conceptual illustration of memory system design principles, even when better-supported alternatives exist

Adoption Level Analysis

Small teams (<20 engineers): Potential fit for solo developers or very small teams wanting local-first agent memory without managed services. The zero-dependency claim and npm install approach are genuine advantages for quick setup. However, the project is early-stage with an anonymous author, unverified benchmarks, and no production case studies. Treat as experimental.

Medium orgs (20–200 engineers): Does not fit. Medium teams need memory systems with proven reliability, multi-user support, access controls, and ideally SOC 2 or similar compliance. Hippo’s SQLite backend is single-process, and the project offers no team-sharing or access control features. Better alternatives (Mem0, Honcho, Weaviate Engram) exist for this scale.

Enterprise (200+ engineers): Does not fit. No enterprise features, compliance certifications, SLAs, or organizational access controls. Not designed for this use case.

Alternatives

AlternativeKey DifferencePrefer when…
Beads (bd)Graph issue tracker with SQLite/Dolt backend, task-centric structureYou want structured memory tied to tasks/issues rather than semantic search
HonchoDialectic user modeling, peer-entity architecture, Python-firstYou need user-centric memory with relationship modeling, not project-centric
Weaviate EngramManaged cloud, vector search on Weaviate, MCP integrationYou already use Weaviate and want managed memory infrastructure
CLAUDE.md / MEMORY.mdFile-based, zero operational overhead, natively understood by Claude CodeYou want the simplest possible persistent context with no external tooling
Mem0Multi-backend (19 vector stores), graph memory, cloud + self-host, SOC 2You need production-ready memory with compliance and don’t want to maintain the stack
Letta (MemGPT)Self-editing three-tier memory (core/recall/archival), agent-managedYou want the agent to manage its own memory autonomously with no CLI intervention
LangMem SDKLangChain-native, episodic/semantic/procedural memory types, open-sourceYou use LangChain/LangGraph and want integrated memory without a separate tool

Evidence & Sources

Notes & Caveats

  • Anonymous author with no prior track record: “kitfunso” has no public profile or prior open-source work visible. The project launched via a Show HN post. This does not disqualify the work but substantially raises the bar for trusting unverified claims.
  • Headline benchmark is unverifiable: The claim of reducing agent trap rate from 78% to 14% has no published methodology, test set, or comparison baseline. It cannot be credited as a benchmark. The metric “trap rate” does not appear in standard agent memory evaluation literature.
  • License not clearly stated: The GitHub repository does not prominently display a license. Before using in production, verify licensing terms — if no license is specified, the code is technically all-rights-reserved by default under copyright law.
  • Naming collision risk: The project shares conceptual branding with HippoRAG (NeurIPS 2024), a more rigorous peer-reviewed system. Search results may conflate the two. Hippo Memory did not cite HippoRAG; HN commenters flagged this as an oversight.
  • Wall-clock decay may not match agent usage patterns: An HN commenter correctly noted that exponential decay tied to wall-clock time disadvantages intermittent agents (agents that run in bursts with long idle periods). Memories may decay to stale status between agent runs even when the underlying context remains valid. The project author acknowledged this and stated they were implementing fixes in v0.10.0.
  • Optional embeddings create two-class users: Installing without @xenova/transformers gives BM25-only search (good for exact-term queries, poor for semantic similarity). Installing with embeddings adds ~400MB and downloads transformer models. The user experience diverges significantly between these two configurations.
  • Multi-agent memory sharing is early-stage: The peering and transfer scoring features are described but no production evidence of multi-agent setups exists. The scope and reliability of cross-project memory transfer are unknown.
  • Node.js 22.5+ requirement: This is a relatively recent Node.js version requirement. Many production environments and CI systems may be running older LTS versions (18.x, 20.x). This constraint limits adoption without disclosure.

Related