Technical proof · updated 2026-08-23

Measurable performance at finding and using the right context

Good answers across sessions require both the right evidence and a reader that uses it. LoCoMo measures the first step; BEAM-100K measures the end-to-end pipeline. We publish the numbers and configuration for each run.

  • 96.58% evidence recall@10 on LoCoMo. +2.1 pp vs published open-source leader (94.5%)
  • 80.8% end-to-end binary accuracy on BEAM-100K (323/400 correct across 20 conversations; reader GLM-5.3 Flash (Ox Alpha); judge GLM-5.2; effort=max; run 2026-08-23; ~44% slower than default effort; this benchmark setup is not the chat default)
  • LoCoMo retrieval: ~0.84s per query on CPU · 0 LLM calls in ranking

ValorBrain · LoCoMo

96.58%

evidence recall@10 · 1,986 questions

Published open-source leader

94.5%

pmb · recall@10

BEAM-100K · end-to-end

80.8%

323/400 correct · 20 conversations

Ranking latency

~0.84s

CPU · 0 LLM in ranking

LoCoMo: +2.1 pp vs pmb. Canonical run: 2026-07-29; 1,918/1,986 hits (96.58%).

BEAM-100K, run 2026-08-23: GLM-5.3 Flash (Ox Alpha), effort=max, judge GLM-5.2. This setting was about 44% slower than default effort. It is an end-to-end benchmark result, not the chat default or independent validation.

What we measure and what we do not

We measure

  • Long-horizon conversational evidence retrieval (LoCoMo recall@10)
  • End-to-end multi-session memory answers (BEAM-100K binary accuracy)
  • Ranking without calling an LLM in the retrieval path

Do not confuse with

  • ·Multi-tenant isolation or role-based permission filtering
  • ·Team handoffs, knowledge gaps, or veracity/origin badges
  • ·Production Ask citation fidelity or latency under live tenant workloads

LoCoMo measures retrieval; BEAM-100K measures end-to-end answers on synthetic conversations. Organizational memory (permission, origin, multi-person + multi-agent) is product capability outside both scores.

Why this matters for the product

In plain language: when someone asks “what did we decide?”, the system must surface the right evidence without inventing and without spending LLM tokens just to rank. That is the foundation; permission, origin, and shared memory for people and agents sit on top.

Fast, cheap ranking

CPU, sub-second, zero LLM calls in the ranking step. Predictable cost.

Private ranking path

The query need not go to a third-party model just to order results.

EN + PT-BR

Bilingual stopwords and multilingual embeddings for real workspaces.

Product foundation

The same core powers Ask, MCP, and organizational memory, with permission and origin on top.

Per-conversation LoCoMo results

ConversationDocumentsHitsQuestionsrecall@10
conv-261919119996%
conv-301910210597.1%
conv-413218819397.4%
conv-422925226096.9%
conv-432923624297.5%
conv-442815415897.5%
conv-473118019094.7%
conv-483023023996.2%
conv-492518719695.4%
conv-503019820497.1%
Total—1,9181,98696.58%

Methodology (honest and reproducible)

  • Per-conversation isolation: clean collection, query, purge.
  • Evidence recall@10 protocol aligned with published LoCoMo; hybrid retrieval plus content-graph reordering (engineering detail lives in the engine repo).
  • Full BEAM-100K split, run 2026-08-23: 323/400 correct answers across 20 conversations. Reader GLM-5.3 Flash (Ox Alpha), effort=max; judge GLM-5.2. Cross-system scores with different readers and judges are indicative only.
  • Numbers on this page come from lib/benchmarks.ts, the SaaS canonical source. We do not invent percentages in the UI.
# Engine (valorbrain repo) — reproducibility
python3 scripts/locomo-benchmark.py \
  --n-conversations 10 --topk 10 --chunk session --mode hybrid

The BEAM-100K run report documents its configuration, latency and failure modes in technical research.

How this compares

Item by item against the alternatives: agent memory, enterprise search and wiki. The table states what each one does and what it does not.

See the comparison

What this saves

The math with your numbers: team size, hours lost looking for context, and what is left after. The assumption stays visible, so you can disagree with it.

Run the math

Organizational memory and context

Prove the retrieval core here. Use the product for permission, origin, and continuity across people and agents.

Start freeFAQ & onboarding