Why Multi-Agent Setups Get Amnesia (and How to Keep Clean State)

What Actually Happens in Production

A dev team built a 3-agent pipeline (Researcher, Writer, Publisher) using CrewAI. By step 6, the agent chat history was 40,000 tokens long. When the Publisher bot was asked to post the draft to WordPress, it hallucinated the post category ID, published to the wrong corporate blog section, and marked the task complete because it lost track of the original user requirements.

Why Writing “Please Don't Do This” in Your Prompt Fails

Attention mechanisms in LLMs degrade as context windows grow (the "Lost in the Middle" phenomenon). When agents exchange intermediate commentary, the signal-to-noise ratio plummets. Asking agents to "remember to check step 1" in their prompts fails because the prompt is overwhelmed by noisy chat history.

The Fix: Putting a Real Lock on the Door

Exogram decouples memory from conversation history. Instead of passing massive text transcripts, each agent writes structured state receipts to an immutable SQLite ledger. When an agent needs to know what happened previously, Exogram fetches only the verified facts via 2-hop graph traversal in 0.07ms.

from exogram import ExogramLedger

ledger = ExogramLedger(session_id="research_task_992")

# Agent 1 records verified output to tamper-proof state
ledger.commit_state(
    entity="article_draft",
    facts={"title": "Q3 Market Analysis", "category_id": 42, "approved": True}
)

# Agent 2 reads clean ground-truth without reading 50 paragraphs of chat
current_state = ledger.get_verified_state("article_draft")
assert current_state["category_id"] == 42 # Guaranteed factual

Frequently Asked Questions