Memory

CoeOS memory rests on two distinctions:

• Write or read. Memory is what the assistant remembers about you, compiled from your conversations, per user. Knowledge is a reference corpus you query (labour law, case law, a large file), shared.
• Open or confidential. What was said in confidence is stored apart, and is never read by an open conversation.

---

1. The levels

Level	Content	Mechanism
User	what the assistant remembers about you, compiled from your conversations	personal wiki, written by the system
Project	the notes and files of a matter (a ZIP upload, or an external folder read live)	files attached to the project
Company	reference corpora	document bases (RAG), read-only

2. Memory split by class

Every memory read goes through a single gate, which receives the class of the turn:

• an open turn only sees open memory;
• a confidential turn sees open and confidential memory;
• data without a known class is never served to an open turn.

The wiki is therefore kept in two parts, stored in separate tables:

• the open wiki, compiled from open conversations;
• the confidential wiki, compiled from confidential conversations. Compilation goes through the box in the confidential class: it only reaches safe models (local, or under DPA and ZDR).

A canary test continuously checks that an open turn never sees confidential memory.

3. The personal wiki (write)

• Per user, no leakage: each note belongs to its user. Nobody sees someone else’s memory.
• Compiled, not raw: a compiler model distils conversations into durable wiki entries (“what to remember”), without duplicating transcripts.
• When: two scheduled passes a day, a compile when a conversation goes idle, and a “Remember now” button on demand.
• Monitored: a health endpoint and an alert flag a stalled compile, so you never end up without memory without knowing it.

4. Finding the right passage

For each question, CoeOS does not paste the whole wiki into the context: it looks for the useful passages.

• Hybrid search: a semantic ranking (bge-m3 embeddings, computed by an embedder on the internal network) and a full-text ranking, merged.
• Sorted by a decision model: the best candidates are scored by a decision model under ZDR, which keeps the ones that really help answer. Measured on 40 questions: the right note comes first 35 times, against 29 without sorting. If sorting fails, the original ranking is kept: it never takes anything away from the turn.
• Only the question goes out to the embedder; the search itself runs in the app.

5. Company knowledge (read)

• Several bases: a registry of RAG bases (Settings › RAG databases), each one can be enabled, tested, and carries an open or confidential label.
• Attached to a project: a project’s conversations query the base linked to it.
• One base, one label: confidential and open are never mixed in one search. A base whose processing goes through a cloud without a DPA is never served to a confidential turn.
• Feeding it: a base is populated by ingesting a corpus (through OdyRAG), then read.

6. How it comes together

On every turn, CoeOS assembles the context with:

• the relevant passages of the user’s wiki, within the limits of its class;
• the project memory, if the project has one;
• the passages of the linked document base, if the class allows it.

The wiki keeps the assistant consistent over time; the base gives it documentary material.

---

In one sentence

The wiki remembers, the base documents, and the class decides what can be read: an open conversation never sees what was said in confidence.