LongMemory 1.0 · Mac · Windows · Linux
Ultrafast memory layer for AIRuns completely offline0 LLM calls
97% token savings.
Compared with sending your AI the whole conversation every time it answers a question (32,264 tokens in our test), LongMemory sends only the notes that matter for that question: ~928 tokens on average. A token is a small piece of a word, and AI is billed by tokens, so that is 97% fewer tokens to pay for on the reading side. In high accuracy mode, which scored 91.5%, it sends ~2,270 tokens per question: still 94% fewer.
Measured on LoCoMo story 1 (the 92.7% piggyback test) · counts only what the AI reads. Token counts
Over 90% accurate.
Two ways.
Tested on LoCoMo, a well-known test where an AI answers questions about long conversations using only what it remembers.
Every message saved exactly as written. No AI notes, no AI calls. One setting turns it on.
Read the full log →Your chatbot’s own reply hands back short, summarized notes worth keeping. No extra AI call.
Read the full log →* Story 1 of LoCoMo only (150 questions), not all ten stories. 91.5% is all ten stories (1,536 questions). Both graded by Claude Sonnet 5.5 with Mem0's published grading instructions, one run each.
96,000 searches at the same instant.
0 errors
One AWS server running 96 copies of LongMemory, each sent 1,000 searches at the same instant. Every one answered.
96 separate LongMemory instances on one machine, 1,000 searches each, median of 3 rounds · 1M tokens saved as one document; 12 ms is the typical search, 99% under 19 ms. AWS run log · 1M-token results
Can run on
edge devices
No GPU, no internet and no LLM calls means memory can live right on the device. Your AI keeps remembering in the field, in the air or on a shop counter, even with no signal at all.
Remember routes, no-go spots and past inspections on board, and keep working when the signal drops.
Runs on Raspberry Pi 4, 5 and Zero 2. Everything it learns stays on the board.
LongMemory builds for Linux on ARM64 (Raspberry Pi 4 and 5, NVIDIA Jetson) and ARMv7 (Raspberry Pi Zero 2), as well as Mac, Windows and Linux PCs · no GPU or internet needed · drones, robots and the other devices above are example uses · images are illustrations.
Extra AI calls.
Most memory products send every message to an AI to decide what to keep, and bill you for it. LongMemory itself is plain, fast code with no AI inside. Save messages exactly as they are, or let the reply your chatbot is already writing hand back the notes worth keeping. Nothing extra to set up or pay for.
584×
faster.
Same questions. Their numbers. Our stopwatch.
How long one search takes. "Typical" means the middle result · Mem0: from its own published results, all 1,540 questions: about 3.1 seconds typically and about 10 seconds for the slowest 1 in 100, through its online service, returning 200 memories per search · LongMemory: the same questions, run here on a Mac laptop (Apple M2) with the program on the same computer, returning 20 memories: 5.2 milliseconds typically and 8.9 milliseconds for the slowest 1 in 100 · the two were not timed on the same machine, so treat the gap as a guide to size, not a lab result
Two steps.
Save it. Find it.
Preference engine.
Tell LongMemory that your wife is vegetarian, that she dislikes spicy food, that you love biryani in the evenings, and that you are off sugar until December. It keeps each one for the right person, and remembers when it applies and when it ends. Later, one question gives your chatbot exactly what is true right now. Your chatbot can pass these along while it writes its normal reply, so there is no extra AI call.