Seeding Agent Memory From Old Company Files Without Importing Junk
A selection method for loading old company files into AI agent memory: score currency and ownership, remove duplicates, and seed small reviewed batches.

To import company files into AI agent memory safely, select records by the decision they support, load only the current version of each fact, and reject anything that is a copy, a draft, or already replaced by a newer document. Start with a short inventory of old files, score each one on currency, ownership and usefulness for one defined role, then load a small reviewed batch before you widen the scope. Old folders hold real history, but raw volume is the main source of contradictions an agent will later repeat with confidence.
Key takeaways
- Choose files by the decision they support, not by the folder they sit in.
- Load only the current version of a fact, and record which version it came from.
- Collapse duplicates into one owned record before anything reaches memory.
- Hold drafts, superseded policies and personal notes until a named owner approves them.
- Seed in small reviewed batches, then test retrieval before adding more material.
Why old file archives mislead agents
An agent does not know which file is newest. When it retrieves content, it sees text that looks relevant, and a three-year-old procedure can read as convincingly as the one approved last month. Shared drives also collect copies that were downloaded, renamed, edited locally and emailed back, so the same rule may exist in five slightly different forms. If you load all of it, the agent will not only mix these versions, it will present one of them as the company position.
The problem is rarely the model. It is the memory you give it. A useful memory holds a small number of facts that someone can name, date by version, and correct when the business changes. Old archives usually contain the opposite: many facts, few owners, and no record of what replaced what.
How do you select which old files to load?
Use four questions for each candidate file. Score each one as yes, unclear or no, and keep only files that pass the first three.
- Does this file answer a question that a defined agent role will actually receive?
- Is there a named owner who can confirm that the content is still in force?
- Is this the latest approved version, and can you tell that from the file itself or its metadata?
- Does the file contain facts that are stable enough to outlive the next quarterly review?
The fourth question is the one teams skip. A customer list from last season or a price table that changes with supplier terms can be useful, but it belongs in a system that is refreshed from its source, not in memory that is treated as permanent. Where a fact changes often, store a pointer to the source of truth instead of copying the value. The same logic applies to any record that depends on a person's judgment, which is better captured as a reviewed note than as a raw transcript.
Also decide what each file is for before loading it. A procedure, a glossary entry, a product specification and a meeting summary behave differently in retrieval. Mixing them in one undifferentiated pool makes it harder for the agent to tell a rule from a discussion about a rule. For guidance on keeping long documents intact, see Chunk Size for AI Knowledge Bases: Keep Policies and Procedures Whole.
How do you remove duplicates and outdated versions?
Filenames are unreliable. Names such as "final", "final2", "revised" and "use this one" show intent, not status. Start by grouping files by subject rather than by name, then compare the content of each group. Near-duplicates are common: the same paragraph with one number changed, or a table pasted into a different template. Where a group contains conflicting versions, the owner decides which one is current, and every other version is marked as superseded and kept out of memory.
When two current sources disagree, do not let the importer choose. Escalate the conflict to the owner, record the decision, and load only the resolved fact. A fuller treatment of duplicate facts is in Superintelligent Memory Deduplication: One Fact, One Owned Record.
Illustrative scenario: one returns policy, four copies
Illustrative scenario: a support team has a returns policy in four places. The original PDF was approved two years ago, a revised Word document was shared with a manager but never signed off, a slide deck from a training session paraphrases the policy, and a support macro contains a shorter version with an older deadline. A naive import loads all four, and the agent answers customers with whichever passage ranks highest for their question.
A selection pass changes the outcome. The approved PDF becomes the single source for the policy. The unsigned draft is held with a note for the owner. The slide deck is excluded because it restates the policy rather than defining it. The macro is flagged for correction because it contradicts the approved rule. The agent then has one current answer, and the support lead has a short list of content to fix in the original workflow.
Checklist before a batch enters memory
Run this checklist on every batch before it is written to memory. A batch that fails any item goes back for review rather than being partly loaded.
- Every file has a named owner and a recorded approval or review date.
- Each fact has a source reference that staff can open and check.
- No superseded, draft or duplicate version is included.
- Personal data, credentials and sensitive records have been removed or excluded, as described in Sensitive Data in AI Agent Memory: What Should Never Be Stored.
- The batch is limited to one role or one workflow, so its effects can be tested.
- A small set of questions with known correct answers has been prepared to test retrieval after loading.
The last item matters most. Loading is only the first step. You need to see whether the agent retrieves the right passage for realistic questions, and whether it cites the current version. Prepare the test questions before the import, not after, so the result cannot be shaped by what the agent happened to return.
What Be Human's managed approach does with this
Be Human designs the memory for each agent around a defined role, so file selection starts from the work the agent must do rather than from the size of the archive. Its managed operations include reviewing what enters memory, recording the source and owner of each fact, running recurring evaluations against known questions, and maintaining the integrations that keep sources current. Permissions and approval boundaries are set so that a person decides when a contested fact changes, and the agent can show where each answer came from.
The approach behind Seeding AI Agent Memory From Company Records in Small Reviewed Batches follows the same logic: small batches, named owners and a test before expansion. Guidance on context selection more broadly is available in [Anthropic's engineering notes on effective context engineering for AI agents](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents), and the [NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework) gives a useful structure for assigning accountability to the people who approve what the system knows.
Keeping imported memory current after the first load
The import is the easiest part to finish and the easiest to forget. Once facts are in memory, they continue to answer questions long after the business has changed. Assign each memory group an owner who reviews it on a fixed schedule, and make that review part of the normal operating calendar rather than an occasional clean-up. When a source document is replaced, the replacement should trigger a review of every memory item that cites the old version, not only the file itself.
Watch for signs of drift. Agents that answer with confident but outdated details, or that cite a source nobody recognizes, are usually relying on memory that was never retired. The practices in How to stop stale context from misleading your AI agent and Document Version Control for an AI Knowledge Base: Cite the Current One help you keep sources and answers aligned over time. If you are not sure which of your old folders would help first, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
How much of an old file archive should I load into agent memory? Load only what supports a defined role and a current decision. Most archives contain large amounts of material that is outdated, duplicated or never used, so a smaller, checked selection will serve the agent better than the full folder. Expand the selection only after retrieval tests show the first batch behaves as expected.
What do I do when two old files contradict each other? Do not let the import process choose between them. Send the conflict to the named owner of the subject, record the decision with its date and source, and load only the resolved fact. Keep the superseded version out of memory so it cannot be retrieved as if it were current.
Can I import emails, chat logs and meeting notes? You can, but treat them as a different kind of record. Conversations are often provisional, incomplete or personal, so they should be summarized and reviewed before they become memory. Remove sensitive details first, and load only the conclusions that an owner has confirmed.
How do I know the import worked? Prepare a set of realistic questions with known correct answers before loading anything, then run them after the batch is in place. Check that the agent returns the current passage and cites its source. If it returns an old version or cannot find the answer, correct the selection or the source before adding more files.
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.


