Three claims, and every one of them is measured

The window is not what the spec sheet says

It breaks at 32K, not at 1M

NoLiMa: needle and distractor share no words with the question. No keyword overlap to latch onto.

18 models. 194,480 calls. Same result.

Wrong context is worse than no context

Three questions that break vector search

Embeddings have a proven upper bound

One vector of size d can only ever express so many different answers.

9 of 18

BEIR datasets where plain keyword search beat every dense retriever tested.

Every signal covers another one's blind spot

Adding keyword search to embeddings cut failed retrievals by nearly half.

Add a reranking pass on top and it drops again, to 1.9%.

There is no single memory. There are four

Same pipeline, same corpus, and the stores land twenty points apart.

Elastic reference build

80 keyword + 80 vector
 → rank fusion
 → cross encoder rerank
time decay + usage boost
per user access control

Load the context you need, when you need it

Skills load only when the agent needs them. Everything else stays a lightweight stub.

Skills as stubs

21-39%

fewer input tokens

Top snippets, not documents

27-34%

fewer input tokens

Agents start contradicting themselves at turn twelve

"Eventually the context window fills up, fidelity drops and token costs climb."

Selective compaction

Keep what matters for the next turn, rather than dropping the oldest.

The context store

Results sit in an in-memory file store and enter context only when needed.

Chats past 30 turns, at up to 40% lower token cost.

Five things to steal, whatever you are building on

Two paths, and five numbers worth watching

Five to watch: recall@10, MRR, tokens per turn, compaction events, and partial hits.