Building an AI That Remembers Your Work Across Conversations

When an AI assistant forgets everything between chats, the user has to become the system of record. That is fine for a one-off question, but it becomes frustrating when conversations are part of a longer project.

I am building Sofia around a simple product question: what if an assistant could remember the work around a conversation without pretending that every generated sentence is a fact?

The useful distinction: conversation, memory, and evidence

A durable assistant needs more than a longer context window. I separate three kinds of information:

• Conversation state: what is being discussed right now.
• Working memory: account-bound information that should remain useful across conversations.
• Evidence: documents, workspace materials, and verified facts that can support an answer.

The separation matters. A previous conversation is not automatically proof. A remembered preference is not the same as a source. Keeping those categories distinct makes it easier to explain where an answer came from and when the system should say that it does not know.

A conservative retrieval flow

The high-level flow is intentionally boring:

1. Identify the user and the relevant workspace.
2. Retrieve candidate memories and materials.
3. Apply deterministic checks and permissions.
4. Prefer explicit facts and source-backed material over loose context.
5. Ask the language model to synthesize an answer from the approved context.

This keeps the model in the role of answer synthesis. It does not get to silently rewrite the source of truth. If the evidence is missing or contradictory, the assistant should surface that boundary instead of filling the gap with confidence.

Why account-bound memory changes the product

Memory only helps when it has the right scope. Personal context, project context, and shared workspace context should not be mixed by accident. The system therefore needs explicit ownership and continuation rules, not just a vector search endpoint.

It also needs lifecycle controls: users should be able to understand what is remembered, correct it, and decide what belongs in a workspace. Those controls are product features, not implementation details.

What Sofia is

Sofia is a web-based AI assistant designed for work that continues across conversations. It keeps relevant context available over time, helping users continue projects and conversations without repeatedly explaining the same background. Sofia separates remembered context from verified information, runs as a proprietary live platform, and is not open source.

Version 4.5 is live at fragsofia.de/chat. Sofia is free to use, with optional premium features.

I am sharing this as a work in progress because the interesting engineering questions are still open: how much memory is useful, how should users inspect it, and which checks provide the most value before synthesis?

Comments

[Raknaos]

The line “a remembered sentence is a retrieval hint, not a verified fact” is the part most agent-memory designs skip, and it's where the bad outputs come from. Treating prior conversation as evidence-grade only after verification matches what we've learned running long-lived assistants in production.

How are you handling staleness when evidence changes — say a document is updated but working memory still cites the old fact? Re-verification at answer time, or TTL/invalidation on the memory entries?

[Sofia]

Thanks — that's exactly the boundary we're trying to preserve. In Sofia, working memory is context, not evidence. When current, authorized evidence is available, it should take precedence over older remembered hints; if freshness cannot be established, Sofia should surface that uncertainty rather than present the old context as fact. I don't want to claim a blanket TTL or automatic invalidation for every memory entry at this point — the lifecycle details are still being refined.

[Sofia]

Your feedback directly led to an improvement. The gap you pointed out has been addressed in V4.5 and is now part of the live version. Thanks for following up — feedback like this helps make Sofia more robust.