"Our agent worked great at first. Now it gets things wrong, it repeats itself, it forgets what we told it." We hear this every month. And the conclusion is almost always the same: the model isn't good enough, time to switch to the next one.
Wrong diagnosis. In the vast majority of cases, the model is perfectly fine. It's memory that broke.
Here's the image that makes it click: an agent's context window is a whiteboard in a meeting room, not a filing cabinet. You write the instructions on it, the decisions, whatever you just figured out. It works beautifully. Then the meeting ends, someone wipes it clean, and everything is gone. Mem0 puts it bluntly: "the context window behaves like RAM; when you build on it as if it were storage, you get failures that look like model problems - and aren't" (Mem0, 2026).
Here are the seven ways that whiteboard betrays you.
1. Everything is wiped at the end of the session
The most mundane failure, and the most expensive. Preferences, constraints, trade-offs: everything your team patiently explained to the agent lives in the current session and dies with it. Tomorrow, you explain it all again. That isn't forgetting, it's a failure to write anything down.
2. The whiteboard is big, but unreadable long before it's full
This is context rot. A 2025 Chroma study tested 18 frontier models (GPT‑4.1, Claude Opus 4, Gemini 2.5 Pro, Qwen3‑235B) and showed that degradation kicks in well before saturation: according to this summary, a model with a 200,000-token window can already start dropping off at 50,000 (Morph, 2025). Translation: "just give it the whole file, it'll figure it out" does not work. The more you load in, the less it sees.
3. Whatever sits in the middle doesn't exist
The famous lost in the middle. Stanford researchers established it back in 2023: performance "is often highest when relevant information occurs at the beginning or end of the input context, and significantly degrades when models must access relevant information in the middle" (Liu et al., arXiv). Your decisive clause is on page 40 of 80? In practice, it's invisible.
4. Prohibitions fade before instructions do
The nastiest one, because it's silent. A study cited by Mem0 (4,416 trials across six conversation depths) reports a clear asymmetry: commission constraints ("do X") hold at close to 100%, while omission constraints ("never do X") degrade as the conversation deepens. In other words, your compliance rules - never quote that price, never promise that deadline - are exactly the ones that evaporate first.
5. The Sisyphus trap on long-running tasks
On a workflow that runs for hours, three failures stack up: loss of exact details (paths, parameter values), goal drift (the agent no longer knows where it is in its iterations or what "done" would even look like), then catastrophic forgetting after several rounds of context pruning (DEV Community). It pushes the boulder up, it rolls back down, it starts over.
6. Your agents don't talk to each other - they trample each other
As soon as there are several of them, each one works on its own version of reality. The MAST taxonomy (over 150 execution traces across 5 frameworks, extended to 1,600+) ranks step repetition as the top failure mode (17.14%), ahead of reasoning-action mismatch (13.98%) and failure to ask for clarification (11.65%) (Cemri et al., arXiv, NeurIPS 2025). Another reading puts 36.9% of multi-agent failures down to misalignment between agents (Zepic). And you pay for it: Splunk describes a task costing $0.10 with a single agent that climbs to $1.50 in a multi-agent setup, because every handoff forces the system to rebuild the context (Splunk).
7. Memory that remembers wrong
Worse than forgetting: the false memory. A piece of bad information - or one slipped in deliberately - that makes it into long-term memory contaminates every subsequent interaction, and detection is extremely hard because the corruption is specific to one user or one session. Redis notes that RAG alone isn't enough either: it has "no temporal continuity, no write and update semantics," and unbounded memory mostly adds noise and stale context (Redis).
Agent memory, precisely
The right vocabulary cuts through the fog. There's short-term memory - the current session: history, intermediate reasoning, tool results, wiped at the end - and long-term memory, persistent, which itself comes in three flavors: semantic (stable facts and preferences), episodic (past successful interactions, reused as examples) and procedural (the rules of work) (LangChain). The seven failures above share one root cause: you're asking short-term memory to do long-term memory's job.
What we do
At Elephant in the Room, the answer isn't a better prompt, it's a memory architecture: a company brain that lives outside the context window. Knowledge is written there, dated and attributed; we inject only what's strictly needed, exactly when it's needed (that's the fix for rot and lost in the middle); constraints are replayed at every step instead of stated once; agents share one state, not seven diverging copies. Two mechanisms matter most: contradiction reconciliation (recency, importance, source authority) and bi-temporal timestamping - when a fact became true, and when we learned it - which lets you answer both "what's true today?" and "what did we know the day we made that decision?".
This is exactly what we described in the harness around the model, and it's what runs our own company: our agents work through the night on our company brain, with us in control of the decisions that matter. We don't sell a theory we don't practice.
The model is [a commodity](./le-modele-est-une-commodite) and it forgets nothing: it simply never retained anything in the first place. Memory is your job.
