Skip to content
be-human.ai
← Back to the journal

Context & memory

Session Memory vs Long-Term Company Memory for AI Agents

Session memory holds a task's working context and then expires. Long-term company memory keeps only approved facts. Learn what to promote and who signs off.

Be Human typographic cover with the title: Session Memory vs Long-Term Company Memory for AI Agents
Conceptual illustration created with AI · Be Human

Session memory is the working context an AI agent holds while it handles one task or conversation, and it should end when that task closes. Long-term company memory is the smaller set of verified facts the business chooses to keep for every future session, such as a policy wording, a process owner or a stable customer record. A fact moves from one to the other only when a named person approves it and the write is recorded.

Key takeaways

  • Session memory is temporary working context. Company memory is reviewed, durable and shared across roles.
  • Promote only facts that will still be true after the current task, and that a named owner can confirm.
  • Promotion needs a write gate, a source reference and an approver. The agent should propose facts, never silently save them as rules.
  • Log every promoted fact with its source, its approver and a review date, so it can be corrected or removed later.
  • Start with one role and one memory scope, then widen only after the first approvals run smoothly.

What session memory actually holds

Session memory covers the current request, the documents the agent opened, draft replies, intermediate tool results and anything a customer said during this interaction. It is useful because it keeps the agent coherent within the task. It is also temporary by design. A long conversation can feel like memory, but a long context window is only the room the model has to read in during one run, as explained in What Is a Context Window in AI? Why Your Agent Forgets. When the run ends, that room empties unless a separate process saves something from it.

Problems start when teams treat session output as knowledge. An agent may write that the customer confirmed an install date when the customer only mentioned it in passing. If that line is saved as a durable fact, the next session treats an unverified statement as settled. The safe default is that session context expires unless someone with authority decides otherwise. Length is not durability, and a detailed session summary is still only a summary of one interaction.

What belongs in long-term company memory

Long-term memory holds facts that describe how the company works and that stay valid across many tasks. Examples include the approved wording of a refund policy, which team owns warranty exceptions, the definitions staff use for each service type, and stable identifiers for customers, jobs or assets. These facts change, but slowly, and each one has a source document or an owner who can confirm it. Entries that describe a customer should also be kept apart from staff opinions about that customer.

Some material should stay out even when it looks useful. Personal data the business has no reason to keep, credentials, payment details and one-off opinions are poor candidates. For the boundary between useful and risky content, see Sensitive Data in AI Agent Memory: What Should Never Be Stored. When in doubt, leave the fact in session and log it as an open question instead.

A decision guide before anything is promoted

Ask these questions before a session fact becomes company memory:

  • Will this still be true after the current task closes?
  • Does a named person own this fact, and can that person confirm it today?
  • Is there a source document, system record or documented staff decision behind it?
  • Would a wrong answer lead a customer or staff member to act incorrectly?
  • Is the fact allowed to be stored under the company's data rules?

If every answer is yes, the fact is a candidate for promotion. If any answer is no, keep it in session or log it as an open question for the owner to resolve. A fact that fails the source or ownership test should never be promoted just because it came up often in conversations. Repetition shows that people talk about something, not that it is correct.

Who approves a promoted fact?

The approver should be the owner of the business area the fact describes, with a named backup. The agent is not the approver, and neither is the model provider, because neither owns the policy or the process. A warranty fact belongs to whoever runs warranty. A scheduling convention belongs to the operations lead. The agent's job is to propose the entry with its source and the session where it came up, and the owner decides whether it enters memory.

This checkpoint is a write gate: the point between what the agent observed and what the agent will remember. The gate has to be simple enough that owners actually use it, with a short proposal, a yes or no decision, and a log line. Memory Write Gates for Superintelligent Agents: Who Saves a Fact covers the design of these checkpoints in more depth. Access rules matter too, because a fact promoted to company level becomes visible to every role allowed to retrieve it, so the approver should also consider who will see the fact.

Illustrative scenario: a warranty question that became a rule

Illustrative scenario, not a real case: a customer asks an agent whether a unit installed by an outside partner is still covered. During the session, the agent learns that partner installations fall under a separate service agreement. It uses this to answer the customer correctly, and the session context then expires.

The agent drafts a proposal for company memory. The draft says that units installed by partners may fall under the partner agreement, that the install date and partner identifier must be confirmed before standard warranty applies, and it attaches the policy document and the session reference. The warranty owner reviews the draft, notices that the partner agreement has changed, corrects the entry to point to the current document, and approves only the corrected version. The log records who approved it, when, and when the source should be checked again.

Notice what did not happen. The agent did not write its own rule from one conversation, and the customer's remark about the install was not stored as a permanent fact about that customer. The durable entry is the policy, not the conversation.

How to set this up, and where Be Human fits

Start with a short plan rather than a full memory system:

  • Name the memory scopes first, such as account, team and company-wide, as described in AI Agent Memory Scopes: Account, Team and Company-Wide Rules.
  • Assign one owner for each fact type, with a backup who can approve when the owner is unavailable.
  • Require that every promoted fact carries a source reference, a session reference and a review date.
  • Keep a change log so that any fact can be traced back, corrected or removed.
  • Review promoted facts on a fixed routine, checking each one against its source and removing what has expired. The process in AI Agent Memory Audit: How to Review What It Remembers gives a workable structure.
  • Track the proposals owners reject. A steady stream of rejections usually means the agent is proposing too much, which is a design problem to fix rather than a reason to lower the gate.

Be Human designs this separation as part of a first engagement: one useful role, the session and company memory it needs, the approval boundary for each kind of write, and an observable success criterion. In managed operations, those boundaries are monitored, permissions are reviewed when people change roles, recurring evaluations check whether retrieved facts are still correct, and incidents such as a wrong promoted fact follow a defined response. The scope stays deliberately narrow so that every memory decision remains visible to the people accountable for it.

If you want to find the first task worth handing to an agent, take the free Be Human AI diagnostic.

Frequently asked questions

Is a long context window the same as memory? A long context window is the amount of text a model can read in one run, so it works as temporary workspace rather than storage. When the run ends, nothing in that space is kept unless a separate process saves it. Durable memory needs a store, rules for what enters it, and a way to correct or remove entries later.

Should the agent save facts on its own? The agent should propose facts, not write them directly into company memory. Each proposal should carry its source and session reference, and an owner decides whether to approve it. Letting the agent write its own durable facts is how one unverified remark becomes a permanent rule.

Who should approve a fact before it becomes company memory? The owner of the business area the fact describes should approve it, and a named backup should cover absences. The agent and the model provider are not approvers, because they do not own the policy or the process. The approver's decision and date should be logged alongside the fact.

How do we remove a promoted fact later? Each promoted fact should have an identifier, a source and a review date, so it can be corrected, expired or deleted. When the underlying source changes, the owner updates or removes the entry and the change is logged. Routine audits catch facts that no longer match their sources before they mislead anyone.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →