§2.1 · Research thread · active · proprietary

hanumemAI

Madhu

Kansas City · Independent · since 24 September 2026 · updated 29 September 2026

Abstract

Long-term memory for an assistant: the highest published score on LoCoMo, reading about a fifth of the context of the systems next to it. The scores and the context each one read are public. The library, its design and its run log are not. The subject itself is taught from zero in the book AI Agent Memory: Novice to Ninja. Other systems' figures are as their vendors published them, each with its own answering model and judge. None is reproduced here, so the comparisons are context and not a ranking.

1 · Article 2 · Paper The book Request access

Keywords: memory.

1Measurements

**Table 1.** Every headline number this thread claims, with the conditions it was measured under and the file it is read from.
measure	value	conditions	source
LoCoMo, all 1,540 questions	96.1	1,533 context tokens per question; 91.6 on the same answers under a stricter judge. Nearest published: Agent Zero 93.6 at 45,800 tokens, Mem0 92.5 at 7,000	`memory research log (private): STATUS.md §1`
LongMemEval-S, all 500 questions	94.0	2,428 tokens per question; 91.8 under a second judge. Not first: 96.2 is the highest published	`memory research log (private): STATUS.md §1`
BEAM-1M, all 700 questions, mean rubric	68.5	2,716 tokens per question; pass rate 73.4. Ahead of Mem0 (64.1 at 6,700 tokens), behind Hindsight (73.9 at 23,900)	`memory research log (private): STATUS.md §1`
Logged runs	162	23 changes kept, 40 discarded	`memory research log (private): results.tsv`

§Where the work lives

§Related writing and lessons

How do you actually measure a RAG system? (writing)
What is Context Precision? (writing)
There's an LLM Running in Your Browser Right Now — No Internet, No API, No Credits (writing)
What is Faithfulness? (writing)

Ask this thread

← All threads§2.2 Documents that cite themselves →