THE BUSINESS AI PLAYBOOK
AI agent memory and company knowledge: what should persist?
Design AI memory that preserves useful context without confusing old conversations with current policy, customer records or permission to take action.
Discuss with Oliver and his team 
AI memory should preserve approved context that helps future work while keeping current business records and access permissions authoritative. Separate company knowledge, customer facts, task state and personal preferences. Give stored information a source, owner and expiry rule. The aim is continuity without allowing an outdated conversation, mistaken summary or private note to become an unquestioned instruction for the next task.
What to take away
- Separate stable knowledge from live records and temporary task state.
- Store provenance and review rules with remembered information.
- Apply access controls when information is retrieved, not only when it is stored.
Name the types of information
Company knowledge includes approved processes, terminology and product guidance. Live business facts include balances, inventory and contract status. Task state records what has already happened in a particular case. Preferences include an approved writing style or a user’s chosen report format. These categories need different update and retention rules.
A single memory bucket blurs those distinctions. The sentence “this customer prefers email” may be a useful preference, while “this customer owes nothing” is a time-sensitive financial fact. The latter should come from the current system of record when needed, rather than a remembered conversation.
Create a memory admission policy
Ask what is worth storing, who may approve it and how it can be corrected. A good memory item includes its source, the relevant person or account, creation time, owner and validity conditions. Avoid saving unsupported inferences about people or treating a one-time exception as company policy.
In a hypothetical customer success workflow, an account manager approves a note that the customer wants quarterly reviews to include a specific usage report. The preference is linked to the account and its source conversation. When the contact changes roles, the next owner reviews that preference instead of assuming it applies forever.
Retrieve narrowly and reconcile conflicts
Bring in the information needed for the current task, with source links and dates. Do not include every past conversation just because it is available. A focused context packet is easier to inspect and reduces the chance that unrelated material shapes the answer. Access should be evaluated for the current user and purpose.
When two sources conflict, use an explicit precedence rule and surface unresolved differences. Current approved policy may supersede a historical example; the live CRM may supersede an old summary. The agent should not merge contradictory facts into a confident story that obscures the disagreement.
Make forgetting and correction operational
Provide a way to review, correct and remove stored context. Decide how changes propagate to search indexes, summaries and caches. Deleting a source document while leaving an accessible summary can defeat the intent of the deletion. Coordinate retention choices with the people responsible for privacy, security and business records.
Copy-ready memory review brief: “List the stored context used for this task, its source and age. Identify facts that require a live lookup. Flag conflicts and expired items. Suggest changes for the owner to approve; do not convert this conversation into permanent policy automatically.” Use the review when employees change roles or workflows change.
FROM READING TO ACTION
Your first steps
- Separate policy, live facts, task state and preferences.
- Require source references and an owner for durable memory.
- Set expiry or review rules for time-sensitive context.
- Test retrieval under different employee permissions.
- Document how corrections and deletions reach indexes and summaries.
What to measure
| Measure | How to track it |
|---|---|
| Stale-context incidents | Track outputs affected by outdated or superseded stored information. |
| Source coverage | Measure the share of durable memory items with an identifiable source and owner. |
| Context correction time | Measure the time from an approved correction to its effective removal or update across retrieval paths. |
Common questions
Should an agent remember every conversation?
No. Store only context that serves a defined purpose and is appropriate for future use. A transcript may contain temporary assumptions, sensitive information or instructions that should never become policy. Deliberate memory is more manageable than indiscriminate accumulation.
Is a vector database the same as memory?
It can help retrieve stored material, but it does not by itself decide what should be retained, who may access it or which source is current. Those are product and governance decisions that need explicit rules around the storage technology.
Sources & further reading
These sources provide background. The workflow designs are Palnau Systems implementation guidance; examples are illustrative, not reported client results. Product capabilities and requirements can change.
Prepared with AI assistance for business owners and managers. For a recommendation based on your actual systems and requirements, work directly with Oliver and his team.
