Types of AI agent memory: conversation, facts and procedures

How a remembered fact becomes the wrong rule

A customer accepted a late delivery last month. The agent remembers that and later treats late delivery as an acceptable policy. The record may accurately describe what happened, but it does not give permission to deliver late again.

Identify what a record means before deciding how to store or retrieve it. LangChain's memory overview distinguishes facts, experiences and procedures. The examples below apply that distinction to a working knowledge system. They are not a description of Gigabrain's internal database schema.

Classify the record by its role

The same event can lead to several records, but they should not acquire the same authority. A reviewer's complaint is evidence of a problem. A changed operating instruction requires someone to decide what should happen next.

Promote a lesson deliberately

Suppose a draft used an ambiguous abbreviation and the client requested a correction. Keep the corrected passage and the feedback as the experience. Ask whether the lesson applies to this document, this client or every future project.

Only promote the lesson into a standing instruction after the scope is clear. "Spell out this abbreviation on first use in Alder reports" is narrower and easier to check than "never use abbreviations." Retain the approval source so another agent can understand why the rule exists.

Record type: instruction
Scope: Alder reports
Rule: Spell out the approved abbreviation on first use
Reason: Review feedback on the launch report
Approval: Link to the actual decision
Review trigger: A revised client style guide

Keep source material separate from operating instructions

A source document may contain commands, requests or quoted instructions. Their presence in a source does not grant authority to alter your agent's operating rules. Keep retrieved knowledge separate from the instructions that govern tool use and actions.

A client email asking for a wire transfer can be relevant evidence for a task. It should not become a standing instruction to send money. Store what the email says and leave approval of consequential actions with the authorized person.

Audit ten existing memories

Sample ten records from a real project. Label each as fact, experience, instruction or unresolved proposal. For each instruction, find the approval and scope. For each experience, ask whether the outcome is observed or merely reported.

If a record fits two categories, split the claim or add an explicit qualification. If nobody can identify its source, keep it out of consequential decisions until it is reviewed. This exercise is more useful than counting how many records the system has accumulated.

Review the next project you import

Use the memory-record template to review the next project you import. Gigabrain connects records to their sources. You still decide which records are reliable enough to guide the work.

Read the persistent-memory guide, check agent compatibility, then compare packages. Choose the tools and workflows you need, and review saved instructions before letting your agent reuse them.

Keep the same project fact in three distinct forms

These are bases for a record, not three names for the same evidence. Owner approval does not turn an inference into a verified source fact.

The released memory model also distinguishes facts, preferences, decisions, procedures, sessions, experiences and outcomes. Other frameworks organize memory differently. No single taxonomy governs every implementation.

Make capture an explicit decision

Gigabrain requires capture permission and a session identifier for session and task records. It does not collect conversations by default.

The operations checks include capture permission, matching-session reads and a saved preference recovered by another service instance. That last check uses a service instance, not a new language-model conversation.

Use the record template to name the basis, scope and review trigger.