Company memory does not require copying every file into a new database. A safer design indexes references and small approved extracts, then hydrates the selected authoritative source on demand. This preserves ownership where the record lives, reduces stale replicas, and makes access decisions before confidential content enters the retrieval path.
Map canonical sources before building an index
Start with business domains rather than shared drives. List the system that owns pricing, customers, policies, contracts, delivery records, and decisions. For each domain, name an accountable owner, allowed entities, sensitivity, freshness need, and the field or event that proves the record is current.
Create a source map containing stable identifiers and approved retrieval routes. The map should say where a record can be found, who may request it, which version behavior applies, and what to do when it is missing. It should not contain copied secrets or broad document bodies merely for convenience.
Build lightweight index entries from metadata needed for discovery: source reference, title or topic, entity, owner, effective date, sensitivity, content hash, and perhaps a short approved extract. The architecture described in company memory beyond vector search provides relevant public context without proving scale or performance for another environment.
Restricted repositories can remain outside ordinary indexing. Their source map entries may reveal only that a protected source exists and which controlled process can evaluate access. This avoids turning the search layer into an uncontrolled copy of legal, financial, customer, or personnel files.
Hydrate one selected source under policy
A query first receives an authenticated identity and operating context. Deterministic rules narrow eligible entities and sensitivities before lexical or semantic ranking. The retrieval stage returns references, not blanket content. Only selected candidates proceed to a permission check and source hydration.
Hydration reads the current authoritative record through an approved connector or local path. Record the expected hash or version so a change between search and read becomes visible. If the source no longer matches, the system should re-rank, request review, or refuse instead of answering from the stale extract.
Use bounded caches only when the workflow needs them. Define what may be cached, encryption and tenant boundaries, expiration, deletion, and invalidation after a source change. A cache is not a new owner of the knowledge. It is a temporary operational copy with a recorded purpose.
The trust checklist for confident AI failures is useful when reviewing stale or unsupported answers. It does not certify that a particular connector, cache, or permission model is secure.
Prove that references stay current and bounded
Test the difficult cases: a file changes after indexing, a source is deleted, a user changes role, one entity has a similar policy, the selected path is unavailable, and an ordinary query resembles a restricted topic. Define the allowed evidence and expected failure behavior before running the test.
Measure retrievals that hydrate current approved content, stale-reference detection, denied out-of-scope requests, cache expiry, conflict routing, and the ability to reconstruct the source used. Do not treat the number of indexed files as a quality metric. A smaller map with reliable hydration can be more useful than a duplicated drive.
The verified completion memory release publicly documents mechanisms for retaining evidence after a task. It does not establish universal recall quality or customer outcomes.
The implementation deliverable should include a source map, index schema, hydration contract, access decision, cache policy, deletion behavior, and regression set. It should also name sources deliberately excluded from search. This turns not copying everything into an explicit architecture rather than an informal promise.
Frequently Asked Questions
Each canonical record remains under its source owner's control, including its access rules, version behavior, and conditions for reading it into company memory.
Keep the source reference, approved extract, content hash, access scope, and hydration receipt. Those fields should tie each retrieval to the current approved record without duplicating the entire repository.
A full copy creates stale replicas, broadens exposure to confidential files, and turns the retrieval layer into another uncontrolled owner of business data.
Include a source map, index schema, on-demand hydration contract, access rules, bounded cache policy, deletion path, excluded sources, and regression tests.
AI4SALE has built source-backed company-memory and bounded retrieval mechanisms and publishes their contracts, without claiming unmeasured business results. Teams can use the exact planned AI4SALE enterprise AI search service after its publication gate to design and test this file-first approach.