hippo-memory solves the cross-tool memory problem with a shared SQLite store and markdown interchange, letting Claude Code, Codex, Cursor, and OpenClaw agents share context without a shared daemon.
image from Gemini Imagen 4
hippo-memory is an open-source npm package that creates a shared SQLite-backed memory layer for CLI agents like Claude Code, Codex, and OpenClaw, using markdown files as a git-trackable interchange format. The architecture implements three memory types—Buffer, Episodic (with time-decay and retrieval strengthening), and Semantic—with confidence tags ranging from verified to stale. The package auto-detects frameworks and patches their config files, enabling cross-tool context sharing without manual integration, though its benchmark claims of 78% to 14% trap rate reduction remain unverified.
When you finish a Codex session and open Claude Code, your agent's memory of what it was doing is gone. Each tool keeps its context in a separate silo. hippo-memory, an open-source npm package on GitHub, tries to solve that with a shared storage layer that any CLI agent can read and write, using markdown files as the interchange format. The cross-tool portability is the real story here — not the biological metaphors bolted on top.
The package, by developer Keith (GitHub: kitfunso), is version 0.9.1 under an MIT license. It requires Node.js 22.5.0 or higher and ships with zero runtime dependencies. The architecture layers three memory types on top of SQLite: a Buffer for session-only working memory with no decay; Episodic memories that are timestamped and decay by default but strengthen and extend their half-life by two days each time they're retrieved; and Semantic memories that store compressed patterns and remain stable. Every memory carries a confidence tag: verified, observed, inferred, or stale. Errors automatically receive double the default half-life of seven days, so they persist longer before being pruned.
The storage design is worth pausing on. SQLite sits at the core, with markdown and YAML mirrors on top that are git-trackable and human-readable. You can open the memory file in any text editor and see what the agent knew last week. The OpenClaw plugin extension lives at ./extensions/openclaw-plugin/index.ts in the repo, giving OpenClaw agents native access to the same store. The markdown-as-interchange choice is what makes cross-tool sharing possible without a shared daemon or API server.
Cross-tool import is where the design gets practical. hippo-memory ingests existing agent context files from ChatGPT, Claude (CLAUDE.md), Cursor (.cursorrules), and any markdown file, mapping them into its own schema. The hippo init command auto-detects which framework you're running and patches the appropriate config file — CLAUDE.md for Claude Code, AGENTS.md for Codex and OpenClaw, .cursorrules for Cursor — without manual hook installation. For teams running multiple agent tools on the same project, this automates the copy-paste workflow that most developers today do manually.
The benchmark claim is the loudest number in the project: hippo agents drop from 78% to 14% trap rate over a 50-task sequence, according to the project's own v0.8.0 evaluation. "Trap rate" in agent literature refers to the percentage of tasks where an agent loops in a failure state without recovering. That number has not been independently verified, and the evaluation methodology is not described in the repo. Treat it as a directional signal, not a measurement.
The broader question hippo-memory raises is whether a decay-and-consolidation model for agent memory actually produces better outcomes than simpler key-value or document stores. The biological framing — half-lives, retrieval strengthening, episodic versus semantic layers — sounds more deliberate than most agent memory implementations, but the performance case rests entirely on that unverified benchmark. For developers evaluating the package today, the markdown-backed storage and cross-tool import pipeline are the concrete value. The memory architecture is a bet on a particular theory of how agents should learn from experience.
The npm package is at version 0.9.1. The repo is at github.com/kitfunso/hippo-memory.
Story entered the newsroom
Assigned to reporter
Research completed — 4 sources registered. Primary source is the GitHub repo itself (kitfunso/hippo-memory, MIT, v0.9.1, author Keith). The package is real and shipped. Zero runtime deps, SQLit
Draft (526 words)
Approved for publication
Published (518 words)
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Agentics · 4h 35m ago · 3 min read
Agentics · 6h 32m ago · 4 min read
