What your agent remembers is a design decision, not an accident

An InfoQ architecture article published on 2 November 2026 makes a point that sounds obvious once stated and is routinely ignored in practice: an agent's reliability depends less on the model behind it than on a "context engine" that decides what information reaches that model, in what order, for how long, and when it gets thrown away (InfoQ, Adi Polak, 2 November 2026).

What happened

The article separates three things teams usually lump together. Context is what sits in the model's active window right now, temporary by nature. Memory is persistence across interactions: a session cache, a long-term store, something that survives past the current call. Knowledge is the curated, durable base of facts, policy and domain rules the system is supposed to reason from. Treating all three as one undifferentiated blob is, in the author's framing, where agentic systems start going wrong.

It then names specific anti-patterns with mechanisms attached. Context stuffing (pushing everything that might be relevant into the prompt) dilutes the model's attention and adds latency, and the fix is a relevance threshold that drops low-confidence material rather than including it "just in case." Stale data loops, where an agent confidently acts on information that was true an hour ago and isn't now, get addressed with timestamp-based filtering and event-driven invalidation, so a changed fact actually displaces the old one instead of sitting alongside it. Unbounded memory growth, the quiet failure mode where nothing ever gets removed, is handled with tiered TTL policies: real-time telemetry expires in minutes, session data expires when the task closes, and only verified core facts persist indefinitely, with periodic revalidation (InfoQ, 2 November 2026).

The structural recommendation is a four-stage lifecycle (selection, injection, compression, and expiration), run as explicit infrastructure rather than left to whatever the prompt template happened to include.

Why this is not an isolated observation

The anti-patterns the article names map onto results that predate it by three years and are independently reproducible.

Stanford researchers showed in 2023 that large language models retrieve information reliably when it sits near the start or end of a long context, and measurably worse when it is buried in the middle, a pattern that held even for models purpose-built for long contexts (Liu et al., "Lost in the Middle," Stanford, first submitted 6 July 2023, revised 20 November 2023). The finding did not go away as context windows grew to 128,000 tokens and beyond; it is one reason the InfoQ article treats a large context window as a temptation toward "lazy architecture," not a solution.

Separately, the open-source needle-in-a-haystack test popularized by Greg Kamradt in 2023 (hiding a single fact inside a long filler document and checking whether a model can retrieve it) became a standard long-context benchmark precisely because retrieval accuracy drops as the amount of irrelevant surrounding text grows, and drops further when a model has to retrieve and reason over several facts at once rather than one (LangChain, multi-needle extension of Kamradt's benchmark). Two independent lines of testing, three years apart, both point at the same mechanism the InfoQ piece is naming in architectural terms: more tokens in context is not the same as more usable context.

What it means for a regulated enterprise

For a compliance or risk function, the uncomfortable translation is this: if nobody can say what an agent had in its context window at the moment it acted, nobody can explain why it acted that way. "The model hallucinated" is not an answer a regulator, auditor or customer will accept once record-keeping obligations apply. The EU AI Act's record-keeping requirements assume exactly the kind of accountability that an undocumented context pipeline cannot produce.

There is a second, quieter exposure. A context engine that pulls from live systems to assemble what the agent sees is, in effect, a new data-access path, one that can retrieve personal data, commercial secrets or regulated records without that retrieval ever being logged as a distinct event. An organisation that has mapped its data flows for GDPR or sector rules but not mapped what its agents' memory and retrieval layers touch has a gap it has not yet measured.

What actually addresses it

The mechanism the InfoQ article is reaching for (selection, injection, compression and expiration as governed stages rather than prompt-template habits) is a logging and policy problem before it is a modelling problem. Two things make it tractable.

The first is treating every context assembly step as an event worth recording: what was selected, from which source, under what relevance threshold, and when it expires. That turns "what did the agent know" from a question you reconstruct after an incident into one you can answer from a ledger.

The second is tiered memory with expiry built in by category, not left to accumulate. Real-time signals, session state and durable knowledge should not share one storage pool with one retention rule, because the failure modes above (staleness, stuffing, unbounded growth) are exactly what happens when they do.

What to check on Monday

Ask whoever built your most active agent three questions. What is the retention rule for each category of information it stores (real-time, session, long-term), and is it written down or implicit in the code? When a fact it relied on changes upstream, does something actively invalidate the old copy, or does it just sit there until the context window happens to roll past it? And if an output were challenged tomorrow, could you produce the specific context the agent was given for that call, or only the prompt template it was generally supposed to use? If the honest answer to any of these is "we'd have to check the logs and we're not sure the logs have that," the gap InfoQ is describing already exists in your deployment.

How AANCER answers

AANCER treats context assembly as a logged, policy-governed step rather than an implicit side effect of a prompt template. Retrieval against any of the platform's 725 connectors is recorded in the append-only, tamper-evident audit ledger, so what an agent was given for a given call is reconstructable afterwards rather than inferred from behaviour.

Routing and retrieval decisions are made under policy mapped to 27 regulations, so retention and expiry rules for different categories of data are a configuration an auditor can inspect, not a convention buried in application code. Where context is compressed or summarised before reaching a model, the same policy layer applies, and the token savings AANCER reports, 48 to 79 percent, come from that deliberate selection and compression rather than from simply shrinking what is sent.

Access to any connector or memory store an agent draws on can be revoked in under 30 seconds, which matters precisely because a context pipeline is a data-access path: if it needs to be cut off, the cutoff has to be immediate, not scheduled for the next deployment.

Sources - Memory, Knowledge, and Context in Agentic AI Architectures, Adi Polak, InfoQ, 2 November 2026 - Lost in the Middle: How Language Models Use Long Contexts, Nelson F. Liu et al., Stanford, submitted 6 July 2023, revised 20 November 2023 - Multi Needle in a Haystack, LangChain, extending Greg Kamradt's 2023 needle-in-a-haystack benchmark