# When memory spans years and sites Scale the search, updates, and evidence checks without losing the question. Author: Murali Sid (https://linkedin.com/in/muralisid) Source: https://www.agenticarchitectureskills.com/memory/the-ten-million-token-question Published: 2026-09-13 Updated: 2026-10-15 Licence: CC BY-SA 4.0, https://creativecommons.org/licenses/by-sa/4.0/ Cite as: Murali Sid. When memory spans years and sites. Agentic Architecture Skills. https://www.agenticarchitectureskills.com/memory/the-ten-million-token-question --- An investigation asks whether the same audit issue has appeared across several sites over years. Reading only the latest report misses the pattern. Sending the entire archive into every call creates a different problem. How will the agent find all the relevant evidence and spot corrections? And how will it keep the search affordable? ## Size is only one part of difficulty [#size-is-only-one-part-of-difficulty] Ten million tokens is an example scale, not a threshold at which one architecture becomes mandatory. Several other things can matter as much as total length. These include the number of similar records and the need to join events across time. They also include permission boundaries and the harm of missing one source. [BEAM](https://arxiv.org/abs/2510.27246) was first published in 2025 and revised in February 2026. It tests several memory abilities on generated conversations of up to ten million tokens. It is evidence that long histories deserve explicit testing, not proof of performance on a company's operational archive. Reviewed 18 September 2026. ## Scope the search, then challenge the scope [#scope-the-search-then-challenge-the-scope] Start with the site, asset class, period, and issue under investigation. Use structured filters and suitable retrieval to assemble candidate evidence. Then test whether those filters excluded related sites, earlier terminology, or an event filed under a different identifier. For a systemic audit investigation, a match at one site is a lead for further collection. Ask what evidence would support or contradict the proposed pattern. A loop should close a specific gap rather than keep searching without a stopping rule. ## Keep a path from summaries to originals [#keep-a-path-from-summaries-to-originals] A site-level summary can help select a year or case to inspect. Preserve links to the underlying records and their versions. A summary of summaries should not become the only surviving account of events. Use exact queries when an official count or status is requested. Approximate similarity retrieval is not automatically an exhaustive population for a company-wide calculation. ## Updates become part of the workload [#updates-become-part-of-the-workload] New reports arrive, identifiers merge, permissions change, and old sources are corrected. Measure the work of refreshing indexes and derivatives, not just the cost of one query against a frozen archive. Partitioning by team or site can help scope retrieval. A cross-site investigation then needs an explicit, authorized path across those boundaries. Avoid assuming a single global index solves access. ## Grow the test before growing the design [#grow-the-test-before-growing-the-design] Test increasing corpus sizes while retaining comparable questions and known supporting evidence. Add near-duplicates, older versions, delayed reports, and relevant cases under different wording. Measure important evidence missed as well as irrelevant evidence returned. Track final decision quality, investigation steps, elapsed time, review burden, and full cost. Keep vendor results separate unless dataset, model, retrieval budget, judge, and scoring rules genuinely match. **Continue the story:** [Read the research findings and what remains open](/memory/what-we-have-measured).