April 6, 2026
Remember This? The AI Doesn’t
Show HN: Hippo, biologically inspired memory for AI agents
An 'AI brain' that forgets on purpose sparks cheers, side-eyes, and séance jokes
TLDR: Hippo is a shared “memory” for AI assistants that saves useful bits across tools and lets old info fade. Commenters split between loving the brainy design, asking for invisible, built‑in plugins and questioning overlap with existing features, while others point to search tools like ccrider and crack séance jokes.
Hippo wants to be the shared memory for your AI sidekicks—saving useful bits across tools like ChatGPT, Claude, and Cursor, while letting old info decay. No database, just Markdown files in your repo, and a daily “sleep” that tidies up. The dev pitch is tidy; the comments? Deliciously messy.
The loudest take: it should be invisible. One critic grumbled “should just run in the background” and argued that “biological memory” shouldn’t be just about time. Others were into the brain vibes—one fan praised the neuroscience analogy, then promptly plugged their own fix: ccrider, a tool to search AI chat logs when they pile up. Cue the debate: do we need a new memory, or better search?
Skeptics asked if Claude already saves project context, and nitpicked naming—why “capture” instead of plain “export”? Meanwhile, the thread’s comic relief arrived with a gem about holding a séance with your past agent, and the crowd ran with it—ghost emojis practically materialized. Another commenter dropped a brainy blog post on “mimicking mechanisms” and linguistic quirks link, adding academic spice.
Bottom line: Hippo’s promise—forget the noise, keep the hard-won lessons—has fans. But the crowd’s split between “this is the AI brain we need” and “nice idea, make it plug-and-play—or just give us great search.”
Key Points
- •Hippo is a shared memory layer for AI agents that stores data in Markdown with YAML frontmatter, is Git-trackable, and has zero runtime dependencies with optional embeddings via @xenova/transformers.
- •hippo init auto-detects agent frameworks, creates buffer/episodic/semantic/conflicts directories, patches hooks (CLAUDE.md, AGENTS.md, .cursorrules), and schedules a daily 6:15am cron job to run hippo learn --git and hippo sleep.
- •Cross-tool import supports ChatGPT JSON exports, Claude’s CLAUDE.md, Cursor’s .cursorrules, and any markdown or text file, with flags for --dry-run, --global, and --tag; duplicates are auto-detected and skipped.
- •Conversation capture extracts decisions, rules, errors, and preferences from transcripts using pattern-based heuristics (no LLM), supporting stdin and file inputs with a dry-run option.
- •Hippo’s memory model uses buffer (session), episodic (timestamped with decay and half-life), and semantic (stable patterns via consolidation); retrieval reinforces memories and errors persist longer. Recall supports budgets.