Unbounded per-agent memory causes long-run degradation and collapse

[Quantum-eon]

Summary

Agent memory/context appears to grow without bound over rounds. On long runs we observed request context grow from ~167k to ~445k tokens, at which point the run enters a cascade of provider 400 (context-length) and 429 errors and effectively collapses. The onset scales with agent count: more agents → earlier collapse. A sliding-window or summarization scheme for agent memory would make long runs viable.

Reproduction

1. Configure a long run (the easiest way, until Issue Proposed changes for fresh builds #2 is fixed, is a long prompt horizon → thousands of rounds) with ~7 agents.
2. Let it run; monitor per-request token counts and provider errors.
3. Around round ~300 (at 7 agents), requests approach provider context limits and 400/429 errors begin cascading; per-round wall time also degrades sharply in the tail (multi-minute twitter/reddit rounds).

Observed

• Context growth from ~167k to ~445k tokens over a long run, then cascading 400/429 failures.
• Degradation threshold tracks agent count, not model family:
  • ~round 300 with 7 agents;
  • ~round 515 with 4 agents;
  • clean, error-free 168-round runs at 10–11 agents (Claude Sonnet 4 and Gemini 2.5 Flash);
  • 18–19-agent runs needed a controlled stop well before their round target, with severe tail slowdown.
• So the practical round ceiling is a function of (agents × rounds), and nothing warns the user where that ceiling is.

Expected

Agent context should stay bounded regardless of round count, so that run length is limited by budget/intent rather than by silent context exhaustion.

Suggested fix

• Bound per-agent memory: sliding window over recent rounds plus periodic LLM summarization of older memory (or importance-weighted retrieval), so per-round prompt size plateaus.
• Fail loudly and early when a request approaches the provider's context limit (log + surface in UI) instead of entering a retry cascade.
• Optionally, document a supported envelope (agents × rounds) until memory is bounded.

Context / disclosure: We are an independent research team auditing multi-agent simulation pipelines. We found this while running a systematic 18-run study on MiroFish-Offline across four LLM families. We plan to publish a reproducible study referencing this issue in roughly 4 weeks, and we're happy to share drafts with maintainers beforehand. Thank you for open-sourcing this project — it made this kind of research possible in the first place.