Case study: 77% of entities lost — then 100% recovered

When an LLM conversation outgrows the context window, platforms summarize it — and lose information. We measured it in our own production multi-agent deployment: standard compaction lost 77% of named entities (people, decisions, tools, dates). Asked about a payment-processor integration discussed 90 minutes earlier, the agent answered "I don't know".

We replaced compaction with a three-layer recall architecture over structured memory: entities survive in full, and the agent retrieves exactly what the current turn needs, on demand.

The result, measured across more than 16,000 messages on two production agent instances: 100% recovery of prior decisions (versus 43% without), while a 46 MB session compresses into 35 kB of working context — a 99.94% reduction in token load. The conclusion of the published paper stands: don't pay for a better model, pay for better memory.

The full methodology, including a review of 27 recent papers from Western, Chinese, Korean and Japanese research groups, is published as an open paper by GLG, a.s. (March 2026) — including Korean and Chinese editions.