Organizational memory should preserve causality.

Video caption

Organizational memory should preserve causality.

New teams inherit a conclusion without its boundary.

Include evidence, dissent, conditions, owner and review trigger.

My rule: Do this because memory is useful only when people can revise it.

#EricFieldNotes

Full written post / accessibility read

When companies add persistent memory to agents, they often celebrate that sessions no longer forget. I want a harder question: can the system explain why a decision was made, what evidence would overturn it and who is allowed to revise it?

Consider a vendor route chosen under one customer contract. The contract changes, but hundreds of tickets still quote the old decision. If the rationale and expiry are missing, leaders face a slow archaeology project when the route fails.

Store source-linked claims separately from the generated summary. Have role agents inject only relevant current constraints and record which version each worker received. A conflict or expiry should surface before the action, not in a postmortem.

Sample a consequential decision and ask a new engineer to reconstruct the source, rejected alternative and validity condition from the packet. If that takes hours or fails, memory volume has not solved organizational learning.

#EricFieldNotes

Four-beat scene transcript

1. Organizational memory should preserve causality.

When companies add persistent memory to agents, they often celebrate that sessions no longer forget. I want a harder question: can the system explain why a decision was made, what evidence would overturn it and who is allowed to revise it?

Visual: A latest-summary file is a weak source of authority.

2. A remembered choice may be expensive to unwind.

Consider a vendor route chosen under one customer contract. The contract changes, but hundreds of tickets still quote the old decision. If the rationale and expiry are missing, leaders face a slow archaeology project when the route fails.

Visual: New teams inherit a conclusion without its boundary.

3. Keep a versioned decision packet.

Store source-linked claims separately from the generated summary. Have role agents inject only relevant current constraints and record which version each worker received. A conflict or expiry should surface before the action, not in a postmortem.

Visual: Include evidence, dissent, conditions, owner and review trigger.

4. Audit recoverability.

Sample a consequential decision and ask a new engineer to reconstruct the source, rejected alternative and validity condition from the packet. If that takes hours or fails, memory volume has not solved organizational learning.

Visual: Do this because memory is useful only when people can revise it.

Research and claim limits

• OpenAI Agents SDK sessions (S53)
• LangGraph persistence (S54)
• OpenAI Agents SDK: handoffs (S170)
• NIST AI RMF 1.0 (S171)

The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.