March 9, 2026
Teach a bot, watch the plot thicken
Show HN: I gave my robot physical memory – it stopped repeating mistakes
Robot gets a memory — and the internet loses its mind
TLDR: An open-source tool lets robots remember past attempts and reuse what worked, showing a 25% success boost in a simple test. Commenters are split between excitement over practical gains and jokes about a “forget” button, while skeptics demand clearer docs and argue it’s clever caching more than true learning.
A hacker says they gave their robot “physical memory,” and the demo claims a simple task got a 25% boost after the bot remembered what worked. The robotmem repo shows a snack-sized API—“learn,” “recall,” even a “forget” button—that stores past attempts and pulls the best ones later. Cue the comment chaos.
The hype squad is ecstatic: “Finally, a robot that learns like a toddler, not a goldfish!” Others love that it runs on a basic computer and reproduces in five minutes. But the skeptics showed up with sirens blaring: multiple commenters argue it’s just “smart lookup,” not real learning, and the term “physical memory” sounds like marketing spin. One recurring gripe: the docs need to explain the comparison chart and what each button does—DANmode literally asks for a plain-text walkthrough.
Then came the comedians: people joked the “forget” feature is the robot version of “delete browser history,” and begged for a “remember my coffee order” demo. Safety worriers asked what happens if the bot recalls the wrong move in a crowded kitchen. Builders want to see more than one test, and whether memories can be shared across robots. Verdict? Big vibes, real promise, and a spicy debate over whether this is genius or just good caching.
Key Points
- •robotmem stores and retrieves robot episode experiences (parameters, trajectories, outcomes) to guide future decisions.
- •A FetchPush experiment reports a success rate improvement from 42% to 67% using robotmem.
- •The experiment is CPU-only and is stated to be reproducible in about five minutes.
- •robotmem provides seven APIs: learn, recall, save_perception, forget, update, start_session, end_session.
- •Recall uses hybrid BM25 plus vector search, with context_filter and spatial_sort; sessions support auto-consolidation and proactive recall.