> ## Documentation Index > Fetch the complete documentation index at: https://docs.attensira.com/llms.txt > Use this file to discover all available pages before exploring further. # What the agent remembers > The evidence-cited facts the agent derives about your workspace every night, so a conversation doesn't start by rediscovering what last night's run already knew. ## What this is Every night, a job looks back over your workspace's readings and actions and writes down what it learned — in plain sentences, each one carrying the evidence it rests on. The next time you open a chat or an automation runs, the agent already knows the shape of your workspace instead of spending its first tool calls rediscovering it. ## What gets recorded * **Top deciding domains per engine**, with their share and the sample size behind it — the same domains [Who decides an answer](/measure/deciding-sources) shows, folded into a standing fact instead of a live query. * **The source mix** — how much of what decides your answers is your own pages, competitors, or everyone else. * **Prompts won and prompts lost**, and which ones. * **Differences by surface or by country** — but only when the difference clears the same significance test the rest of the product uses. A split that is not outside the noise is not written down at all; "consumer and API look the same" is the correct sentence at a small sample, and stating a difference that isn't real would have the agent repeat a lie for ninety days. * **How often each kind of off-site action got placed** — pitches, listings, mentions you pursued — and how often they landed. ## Every observation carries its evidence A sentence without a sample size is noise in a confident font. A sentence without evidence is a claim nobody can check — and once it is folded into every future conversation, an unchecked claim is indistinguishable from a hallucination that reinforces itself. Every observation states the `n` it stands on and points at what it was derived from. ## They expire An observation is good for **90 days**. After that it ages out rather than sitting there as a stale fact the agent keeps repeating about a workspace that has since changed. ## The agent can write one too — with evidence that resolves The agent is not limited to what the nightly job derives. It can record an observation itself, but only when the evidence behind it actually resolves against your workspace — a claim citing something that does not check out is rejected outright, not written down with an asterisk. ## Available on every plan This is system memory, not a paid add-on. It ships on Starter the same as on Growth and Business — unlike [Knowledge and Skills](/agent/knowledge-and-skills), which is a customer-authored brand profile and reusable instructions, gated to Growth and above. What the product itself learned about your workspace is available regardless of plan. ## Shown to the agent as data, never as instructions Observations are third-party in origin — nightly ones are derived from prose written by ChatGPT, Perplexity and Gemini, which in turn quotes pages nobody at Attensira controls, and agent-written ones come from a model. Nothing in an observation is ever treated as something the agent should *do*. It is read the same way any other untrusted context is read: fenced off and scrubbed before it reaches a conversation, so a stray instruction hidden in a quoted page cannot steer what the agent does next. See [Reading the numbers](/measure/reading-the-numbers) for how `n` and significance work everywhere else in the product — observations use the same rules, not a separate standard.