Agents Shouldn't Pass Context. They Should Share Memory

Описание и применение

Многопроцессорные цепочки агентов часто растягивают контекст, увеличивая стоимость и задержку. Решение – использовать совместную память с хранением в трех уровнях (эпизодический, семантический, процедурный) и ограничивать размер передаваемого контента токен‑бюджетом. Это стабилизирует расходы и ускоряет работу.

Задача

Multi‑agent chains accumulate token‑heavy context, driving up cost, latency, and diluting attention.

Результат

Token consumption per round became a fixed budget, eliminating linear growth; cost and latency stabilized across long chains and multi‑session use.

Рабочий процесс

• 1. Agent writes output to an append‑only episodic log (L1).
• 2. Asynchronously project new log entries into the semantic tier (L2) via LLM extraction.
• 3. Agent retrieves a token‑budgeted slice from both L1 (fresh events) and L2 (distilled facts).
• 4. Agent performs its task and appends its result to L1.
• 5. Repeat for each agent in the chain; on new sessions, L2 provides cross‑session beliefs.
• 6. Periodically rebuild L2/L3 from L1 if needed for consistency.

Доказательства

Agents do not hand anything to each other directly. Each agent writes its output to a shared memory layer, and the next agent retrieves only the slice it needs, capped to a token budget.

Источник

The episodic tier (L1) is an append‑only log of everything that happens in a session…

Источник

If Agent A produced forty thousand tokens of investigation and Agent B needs the three thousand tokens of conclusion, B receives three thousand.

Источник