EUMNESIA LABS

AI RESEARCH LABORATORY
EST. 2026 · FIRST PAPER EXPECTED 2026

ACC. NO. 2026-001
EUMNESIA LABS
PERMANENT COLLECTION

The end of AI amnesia. the ink fades — see Exhibit I

EU·MNE·SIA — GREEK EU (GOOD, HEALTHY)
+ MNĒSIS (MEMORY).
THE CLINICAL OPPOSITE OF AMNESIA.

AI models forget. We build attention that keeps memory exact — and reads only what matters — at any context length.

Enter the exhibit ↓ The record ↓
E.01 · THE RECALL WALL

Every system today picks a side.

Efficient models compress the past into a fixed-size memory. Compression is lossy, so they provably lose facts they read moments earlier — the recall wall.

Full attention keeps everything and forgets nothing — and pays a quadratic price for it. Remember, or afford it. Not both.

PLATE 1 · TOKENS 1–36WHAT SURVIVES COMPRESSION
E.02 · WHY NOW

Memory is now the bottleneck.

Context windows are racing toward millions of tokens — and the field is discovering that reading long is not the same as remembering long.

AI agents are now asked to carry weeks of work: codebases, case files, running conversations. A single quietly forgotten fact is enough to break the result.

Compute keeps getting cheaper. Forgetting does not. Memory — kept exactly, read efficiently — has become the bottleneck of useful AI.

E.03 · CONDITION REPORT

The gap is architectural, not incidental.

Condition report — our reproduction of published results. MQAR benchmark, matched capacity.

Full attention

QUADRATIC COST

0.99

Fixed-state models

CHEAP

0.10

Ours — selected, exact attention

IN DEVELOPMENT

with the paper

Forgetting is not a law of nature. It is an engineering choice.
E.04 · METHOD

Keep the memory exact.
Read only what matters.

i.

Exact memory

The past is kept, not compressed away.

ii.

Learned selection

The model learns where to look.

iii.

Sub-quadratic end to end

Including the selection step itself — where others quietly pay the quadratic price.

Our first architecture is in active development. Details arrive with the paper.

E.05 · THE RECORD

What we have actually done.

  1. JUN 2026 COMPLETE

    Reproduced the recall wall — attention ~0.99 vs fixed-state ~0.10 at matched capacity. The gap is architectural.

  2. JUN 2026 COMPLETE

    Verified that content-based sparse attention holds the dense-recall ceiling — its hidden cost is the selection stage itself.

  3. JUL 2026 COMPLETE

    Built our own architecture and validated it across sequence lengths. Full results with the paper.

  4. JUL 2026 COMPLETE

    Controlled budget study: the recall loss at length is recoverable, not fundamental. Full results with the paper.

  5. NOW IN FLIGHT

    Multi-seed confirmation, longer sequences, and head-to-head frontiers against state-of-the-art efficient architectures. Paper expected 2026.

  6. NEXT

    Paper draft

  7. 2026

    First paper

  8. WITH THE PAPER

    Repository opens

Research first.
Claims with receipts.

Every claim ships with open, reproducible experiments. No vendor-run benchmarks. The repository opens with the paper.

Until then we would rather say less than say something we cannot show you.

E.06 · FIELD NOTES

Field notes.

Writings from the lab — plain words, public knowledge, no hype.

F.01 · ESSAYREADING TIME · 3 MIN

Why AI forgets — the recall wall, explained simply

Ask a person to read a long letter and they will miss a few details. Ask most efficient AI models to do the same and they will miss them predictably — not because they are careless, but because of how they are built.

A language model reads text one piece at a time, and as it reads it must keep some record of what came before. There are two classical ways to do this. The first is to keep everything: every word stays available, and when the model needs a fact it looks back at the original. This is what full attention does. It remembers perfectly — and the cost of looking back grows with the square of the length of what it has read. Double the document, quadruple the bill.

The second way is to compress: squeeze the past into a fixed-size summary as you go. This is what fixed-state models do. The cost stays flat no matter how long the text grows. But a fixed-size summary can only hold so much. Once the text carries more facts than the summary has room for, something must be thrown away. The model does not choose badly — it simply cannot keep everything. This is the recall wall, and it is provable, not anecdotal: in controlled tests, models that compress lose facts they read moments earlier, while models that keep everything recall them almost perfectly.

So every system today picks a side: remember everything and pay quadratically, or pay little and forget. But the wall is not a law of nature — it is a consequence of the fixed-size summary. Remove that constraint, keep the past exact, and learn to read only the parts that matter, and the wall stops being load-bearing. That is the work.

APPENDIX A · QUESTIONS

Questions, answered plainly.

Is this a product I can use today?

Not yet — we are research-first. The paper and open repository land in 2026; long-memory models and tooling built on the architecture follow. The experiments behind them are compute-intensive: multi-seed training runs and long-sequence evaluations.

Are you hiring?

Not yet — but write to us.

Can I follow the research?

The repository opens with the paper. Email us to be notified.

Why "Eumnesia"?

Greek eu (good, healthy) + mnēsis (memory) — the clinical opposite of amnesia.

CORRESPONDENCE

anuj@eumnesia.com

Get notified when we publish ↗

FOUNDER

Anuj Dev Singh

AI engineer and researcher — data science, IIT Madras. Works on attention, memory, and long-context recall. Based in Delhi, India.

GitHub ↗Eumnesia Labs on GitHub ↗codewithanuj.com ↗

Eumnesia Labs · founded 2026