Support is not short of memory systems. The CRM has the account, the ticketing system has every prior case, and a retrieval layer pulls what looks relevant into the prompt. The failure is what survives the trip into the model's context: retrieval returns what matched, and a handoff summary paraphrases the identifier it was supposed to carry. A bounded field keeps the history resident and the identifiers exact, and loads your product and policy knowledge as a versioned lens the agent reads directly — so it answers from your current policy, not a stale fine-tune.
Support already has the data. The CRM holds the account, the ticketing system holds every prior case, and a retrieval layer pulls the passages that look relevant into the prompt. Between threads the conversation itself is reset, and what carries over is a generated summary sitting in the handoff notes.
The failure is not missing data, it is what survives the trip into the model's context. Retrieval returns the passages that matched the phrasing, not the ones that mattered, so the account number stays in a record nobody pulled. The handoff summary paraphrases, so an order reference becomes a description of an order reference. The customer, who can see none of this machinery, just knows they are explaining it again.
SCOPE: schematic. Measured recall on your pinned fields comes out of the pilot, against your current retrieval and handoff configuration.
A customer relationship is not one conversation. It is every contact they have ever had with you, and the cost of carrying it is what decides whether you can.
SCOPE: measured on a single unbroken conversation of 1,900,000 tokens across 9,359 turns, 14.5× the base model's trained context, with the resident memory flat from 50,000 tokens through to the final turn and total footprint ~1,370× at 1.9M tokens under a standard cache. The ratio grows with length and is only meaningful at a stated depth. Reproduced on 3 independent conversations. Your recall curve and the configuration that produced it come out of the pilot.
Accounts, cases, and order references pinned verbatim, so they are never paraphrased by a summary or missed by a retriever.
One field per customer carried across every thread, so nothing resets and nobody repeats themselves.
A memory cost per active conversation you set in advance, in place of per-turn retrieval that scales with history and traffic.
Your product and policy knowledge is loaded as a versioned lens, so the agent answers from current policy — and a policy change is a new lens version, not a retraining cycle.
The history stops hitting a ceiling. Before we bounded the read, our own runs died at roughly 285,000 tokens: output collapsed to about a dozen tokens a turn and stopped. A high-contact account accumulates past that. Nothing like that ceiling appears at any depth we have run since.
SCOPE: continuity behavior depends on your handoff design and field budget. A paid pilot returns a measured comparison against your current retrieval and handoff stack. Our published depth results were produced on a deliberately small research model, chosen to stress the memory rather than flatter the output. A larger model abuses it less. Treat them as a floor.
Every prior contact carried as one resident structure that keeps merging as it grows, not reconstructed from handoff notes each time. What was said on the first contact is read at the same resolution as what was said on this one.
The current turn word for word, over a policy block the field cannot overwrite, so the agent has the conversation in front of it without losing the record behind it.
Your support policies and product facts, loaded as a versioned lens the agent reads alongside the customer's history. Diffable and updatable, so what the agent knows changes when you ship a new version — not when you retrain.
Recall on pinned fields, accounts, cases, and references, against your current retrieval stack on the same conversations.
Repeat-ask rate across threads, where a reset and a paraphrased handoff would have forced the customer to start over.
A memory cost per active conversation, next to what per-turn retrieval currently costs you.
Read fidelity against a full-attention reference, so you can see what the compression costs before you choose where to sit.
Policy adherence and knowledge recall from the loaded lens, against your current retrieval or fine-tuned baseline.
SCOPE: this page states direction, not magnitude. Measured numbers for your workload come out of the pilot, with the configuration that produced them.
Bounded memory on its own is not hard. A rolling window gets the same footprint number tomorrow by dropping the oldest turns, and in support, the oldest turns are the account history. What is hard is holding the footprint flat while the whole history stays available, and showing the model is still reading the start of it at depth. That is the result we publish, reproduced across independent conversations, and it is what a pilot reproduces on your workload.
The field is held to a budget you fix before you deploy. It runs on the model you already serve, with no retraining and no recalibration. Patent pending; the implementation stays private.
A scoped four-week pilot: you bring one workload and a memory budget, we return a measured comparison against your current cache: footprint, recall, and the configuration that produced both.