Most AI memory tools save everything and search later — which sounds useful until an agent hits the same bug for the fourth time because nothing told it that lesson mattered. A new open-source project borrows a different idea from actual brains: memories that fade unless something keeps using them.
What it is
Hippo is “a memory layer for AI agents, modeled on the hippocampus,” built with three explicit stages — a session-only buffer, an episodic store where “memories decay by default” and “retrieval strengthens” them, and a semantic store where “repeated episodes compress into stable patterns” during a step the project calls sleep (project site: hippo-memory.com). The GitHub repository (751 stars: kitfunso/hippo-memory) frames the core idea in one line: “The secret to good memory isn’t remembering more. It’s knowing what to forget.”
How the decay actually works
This isn’t a vague metaphor — the mechanics are specific and stated as numbers, not adjectives. Per the project site: “Every memory fades on a 7-day half-life… Each recall extends the half-life… Tag a failure once. It decays slower and resurfaces every time you walk back into that code” — a 2x half-life for tagged errors. The GitHub README states the reward mechanic precisely: “A memory with 5 positive outcomes and 0 negative has a reward factor of ~1.42, making its effective half-life 42% longer. A memory with 0 positive and 3 negative has a factor of ~0.63, decaying nearly twice as fast.” Whether a lesson persists depends on whether it kept turning out to be useful, not just on how recently it was written down.
The project’s own comparison table (hippo-memory.com) claims it’s the only one of four compared tools (against Mem0, “gbrain,” and Zep) that implements both decay-by-default and retrieval-strengthening — a claim worth treating as the project’s own framing of its competitors, not independently verified against those tools’ actual current feature sets.
Benchmark claim, and how much weight to give it
The headline number on the project site is “98.6% R@5 on LongMemEval with the zero-dependency default” — a real, named benchmark (LongMemEval tests long-term conversational memory recall), not an invented internal metric, which is a meaningfully stronger signal than an unnamed “our tests show.” Still worth noting plainly: this is the project’s own reported score on its own site, not yet cross-checked against an independent LongMemEval leaderboard at time of writing.
This isn’t an isolated idea
A separate, peer-reviewed-style paper on decay-based agent memory, FadeMem, published on arXiv by researchers including authors affiliated with Alibaba’s international commerce group and Peking University, describes near-identical reasoning: “current AI systems employ binary retention strategies that preserve everything or lose it entirely,” and proposes “differential decay rates… modulated by semantic relevance, access frequency, and temporal patterns,” reporting “45% storage reduction” on standard long-term-memory benchmarks (arXiv 2601.18642). Hippo and FadeMem are unrelated projects with no evidence of connection between them — the fact that both independently arrived at decay-based memory as the fix for “AI agents forget the wrong things” is itself a signal this is a real direction, not one project’s marketing angle.
Why this matters
Persistent memory is the feature that separates an agent that improves over repeated use from one that starts fresh every session — but “remembers everything forever” turns out to be closer to the problem than the solution, since undifferentiated storage buries the one lesson that mattered under a thousand that didn’t. Whether Hippo’s specific decay curves are well-tuned is an open, unverified question; that the general approach — borrowed directly from how biological memory actually works — is showing up independently in both an open-source tool and separate academic research is the more durable finding here.