How Neumannic memory works
What memory is, what it stores, how to turn it on or off, how to delete it, and the privacy posture that surrounds it. This page is for customers who want to understand the memory feature before they enable it — or who want to verify the policy after they have.
What is memory?
Memory is what lets Neumannic remember context across sessions, so you don't have to re-explain yourself every time you sit down to draft, refine, or review work. It's the difference between starting every conversation from scratch and picking up where you left off — the names of the products you sell, the tone your brand uses, the constraints your team works within, the projects you've already shipped.
Memory in Neumannic comes in three layers. Each layer has its own scope, its own audience, and its own rules for what gets remembered.
This conversation only
Context held for the length of a single conversation. Lets Neumannic refer back to things you said earlier in the same session without losing the thread. Nothing from this layer persists once you close the conversation.
Workspace memory
Shared across everyone in your workspace. Holds the context the whole team relies on: the products you sell, your brand voice, your approval rules, the projects you've already shipped. Any workspace member can read from it; only members with the right permissions can write to it.
Personal memory
Specific to you. Holds your individual preferences and the things you've told Neumannic to remember on your behalf — your writing style, your shortcuts, the way you like to work. Other workspace members cannot see this layer; it stays between you and the assistant.
What does it store?
Memory captures the kind of context you would otherwise have to repeat at the start of every session. Some things are saved because you explicitly ask Neumannic to remember them; others are inferred from your work and surfaced for your review before they're saved.
Common examples of what gets stored:
- Brand-voice notes you've shared with the assistant (“we write in plain English, no jargon”).
- Recurring client preferences (which clients want copy in British English, which products have approval-process constraints).
- Document-specific feedback the assistant referenced in a prior session (the way you wanted a particular landing page to land).
- URLs and documents you explicitly told the assistant to remember as canonical references.
What is never stored automatically, under any circumstance:
- Payment details, card numbers, or billing information.
- Passwords, secrets, API keys, or anything that looks like a credential.
- Personal information you did not explicitly save — government identifiers, contact details, health information.
- Anything that looks like it might belong to a different workspace — the assistant is explicit about workspace boundaries and will refuse to carry context across them.
How do I turn it on or off?
Memory is opt-in. Nothing is captured until you turn it on, and you can turn it off again at any time. Memory has two opt-in surfaces — one for your workspace, and one for your individual account — so you have control at both levels.
The two control surfaces:
Your personal memory
Manage your personal memory at /settings/memory. You can turn it on or off, see what is currently saved, review items the assistant has flagged for your attention, and remove anything you no longer want remembered.
Your workspace memory
Workspace administrators control workspace memory at /admin/workspace/memory. If you are not an administrator and want workspace memory turned on, ask whoever owns your Neumannic workspace to enable it for you.
Memory settings panel
An annotated screenshot of the memory settings panel will appear here. The panel lives at Settings → Memory, where you can turn memory on or off, review pending items before they are saved, and remove anything Neumannic has already remembered for you.
What is the retention policy?
Different kinds of memory items are kept for different lengths of time. The shorter the retention window, the lower the trust we place in the item before you've confirmed it. The table below is the customer-facing version of our retention rules.
Things you explicitly saved: When you tell the assistant to remember something directly, it stays in your memory until you remove it. These are the highest-trust items.
Things our assistant inferred from a conversation: When the assistant notices something worth remembering during a conversation, it flags the item for your confirmation before saving. Confirmed items are kept.
Things our assistant inferred from a URL you shared: URLs you share may contain context the assistant thinks is worth remembering. These items go through a review queue; if you don't confirm within seven days, they're discarded.
Things our assistant inferred from a document: Documents you upload may contain context the assistant thinks is worth remembering. Same as URLs — items go through a review queue and are discarded if not confirmed within seven days.
How do I delete memory?
You can delete a single memory item, delete a whole category of items, or wipe memory entirely. All three paths are available from your account settings. Each path is described below in order, from least invasive to most invasive.
Delete one entry
Open Settings → Memory, find the item in the list, and click the remove control next to it. The item is removed from memory immediately and will not be referenced in any future session.
Delete a category of entries
From the same Settings → Memory panel, you can filter the list by source category (for example: “all entries inferred from URLs”) and remove every item in that category at once. This is the right path when you want to tidy up without starting over.
Delete all memory
The Settings → Memory panel includes a “remove everything” option that wipes every item Neumannic has remembered about you. This is a complete reset — the deletion removes the item from every place it lives, including the search index that the assistant uses to find relevant context.
Heads up
Deleting all memory is irreversible. We cannot restore items you have removed, even immediately afterwards. If you are unsure, consider deleting a single category first to confirm the right path before wiping everything.
How does memory affect what the assistant says?
Memory shapes what the assistant brings up in your sessions, but not all memory items are treated equally. Items you've explicitly saved or directly confirmed are weighted more heavily than items the assistant inferred and you haven't yet looked at. The principle is straightforward: the more directly the item came from you, the more confidently the assistant will reach for it.
The directional rules:
- Items you explicitly told the assistant to remember carry the most weight in retrieval.
- Items set up by a workspace administrator (workspace-wide brand voice, named projects) also carry strong weight — they represent the shared team context.
- Items inferred from your conversations carry moderate weight; the assistant uses them when relevant but treats them as “working memory” that you may want to revise.
- Items inferred from URLs or uploaded documents carry the lowest weight until you confirm them — they sit in the review queue first.
Memory items also fade gradually over time if you don't reference them. An item that hasn't come up in roughly ninety days carries less weight than a recently-confirmed item — not because it gets forgotten, but because the assistant treats older items as weaker context unless you bring them back into focus. If something is important and you want it to stay loud, the simple path is to mention it occasionally.
We don't publish the exact arithmetic for how the weights combine, because the system is intentionally tuned so that prompting the assistant in a particular shape isn't a way to game retrieval. What we can promise is that the directional principles above hold across every tier and every feature: explicit beats inferred, recent beats stale, you beat the assistant's guesses.
How do you keep my memory safe?
Memory is a sensitive surface — if an attacker can quietly drop a false fact into the assistant's memory, the consequences ripple through every future session. We've designed memory with three independent safeguards so a failure in any one layer doesn't compromise your data.
We screen what goes in
Before any item is saved to memory, we screen it for known patterns of attempted manipulation — things that look like instructions to override the assistant's behavior rather than facts about your work. Items that fail the screen are rejected outright; suspicious items go through the review queue rather than landing in memory directly.
We cryptographically sign what is stored
Every item written to memory is signed at write time, and the signature is verified at read time. If an item is tampered with between writing and reading — by any party, including us — the verification fails and the item is discarded rather than served. Tampering is detectable, not silent.
We audit what comes out
When the assistant retrieves items from memory for a session, the retrieval path watches for adversarial patterns — for example, a query that looks designed to extract sensitive items rather than relevant context. Items that match these patterns are dropped from the response and the event is logged for review.
These three layers compose so that a successful attack against Neumannic's memory would have to clear all three — get past the input screen, forge the signature, and evade the retrieval-time pattern check. For broader context on our security posture, see the security overview page.
What about Zero-Data-Retention?
Memory operates under the same Zero-Data-Retention posture as the rest of Neumannic. When the assistant reads from your memory or writes a new item to it, the underlying model provider receives the request, processes it, and returns the result — and then the provider's record of that exchange is gone. Memory writes and memory reads are not an exception to our broader retention policy; they inherit it.
We have set the platform up so that no model provider retains your data after the request completes — including memory writes and reads. This is non-negotiable across every part of Neumannic, every tier, every feature. The items themselves live in our managed memory storage (signed and screened per the previous section); the providers we route through to extract, embed, and retrieve those items do not keep a copy.
For the full Zero-Data-Retention contract — what is and is not stored across the entire platform, how the binding is enforced in code, configuration, and runtime, and the verifiable list of providers we route through — see the privacy overview page.
Still have questions?
If something on this page is unclear, or you want to talk through how memory fits into your specific workflow, email us at privacy@neumannic.com. For the formal version of our data-handling commitments, see the Neumannic Privacy Policy.