Emergence Labs · Active Research

How AI agents coordinate.
And when they shouldn't.

Every design choice in atlaso started as a result we ran. Memory architecture, agent coordination, timing — published with full methodology, judge prompts, and raw run logs.

3Papers·4Judges · n=500·Reproducible·Open source

Read the papers

Published·3 papers

Updated May 2026

memory_benchmark.fig1+9.8 to +14.8pp · 4 judges

Atlaso vs mem0 · 4 judgesfig 01

01·Benchmark Study·May 2026

What memory systems actually do.

A four-judge head-to-head on LongMemEval-S (n=500) with a shared reader. Atlaso beats mem0 by +9.8 to +14.8 percentage points across four independent judges, at 1.20× lower cost per query. Mem0's published 93.4% does not reproduce under matched conditions.

n=5004 judgesShared reader

10 min readRead paper

adaptive_gate.fig1+34% on hard tasks

n=10 seeds · hard taskfig 02

02·Research Paper·Apr 2026

When to start communicating.

An adaptive gate that activates stigmergic coordination only under population-level distress signals outperforms both always-on and always-off baselines by +34% on the hard task across 10 seeds — and the same principle transfers to LLM agent collectives.

10 seedsPre-registeredLLM transfer

8 min readRead paper

growing_field.fig2+11.87pp · 10/10 seeds

10 seeds × 300 tasksfig 03

03·Pilot Study·Apr 2026 · Updated Apr 25

Where the 12-point lift comes from: 94% retrieval, 6% format.

On 10 seeds × 300 tasks, the +11.87pp lift on Qwen3.5-9B decomposes cleanly. Matched-token random ICL: ~0pp. Scope-matched content with the deposit schema stripped: ~94% of the lift. Trained deposit format adds ~6% (+0.70pp, CI [+0.60, +0.80]).

10 seeds × 300 tasksPre-registeredCI computed

12 min readRead paper

More research in progress. Subscribe below for new releases.

How AI agents coordinate.And when they shouldn't.

Published·3 papers

What memory systems actually do.

When to start communicating.

Where the 12-point lift comes from: 94% retrieval, 6% format.

New papers, pilot results, and things we're learning.

How AI agents coordinate.And when they shouldn't.

Published·3 papers

What memory systems actually do.

When to start communicating.

Where the 12-point lift comes from: 94% retrieval, 6% format.

New papers, pilot results, and things we're learning.

How AI agents coordinate.
And when they shouldn't.

How AI agents coordinate.
And when they shouldn't.