- Org memory — durable facts and preferences that apply to the whole organization (e.g. “never mention pricing before a demo is booked”). Org memory is loaded into every conversation automatically. Writing to it is Admin-only. Where: the Org tab at agent.getswan.com/memory.
- User memory — your own personal preferences and notes (e.g. how you like emails written, your calendar habits). Only you can write to your user memory, and it’s loaded into your own conversations. Where: the Personal tab at agent.getswan.com/memory.
- Record memory — context tied to a single record, such as a company, contact or deal (e.g. “this account had a bad support experience in 2024, mention carefully”). Every record has a folder in the brain,
/brain/records/{object}/{slug}/(a company’s is/brain/records/companies/{domain}/), and its memory is theMEMORY.mdfile inside it. The folder appears once something is first saved to it. Swan reads and updates a record’s memory when it’s actively working on that record, rather than loading it into every conversation by default. Reading record folders needs Record read permission and writing needs Record update. Where: for a company, the memory card on its page (https://agent.getswan.com/companies/{companyId}) or the full-page view athttps://agent.getswan.com/companies/{companyId}/memory.
Memory size caps
Each user memory is a single document capped at 2,500 characters of plain text; org memory and each record’s memory allow 10,000 characters. Markdown formatting (headers, bold, bullet points) doesn’t count against the cap, so you’re free to structure a memory for readability without it eating into your budget. When you ask Swan to check a memory’s status, it reports how much capacity is used. Once a memory crosses roughly 70% full, Swan treats consolidation as routine internal maintenance and will trim it before adding more — collapsing redundant notes, dropping stale information, tightening phrasing — so you don’t lose new information to a hard limit. Swan doesn’t notify people or create tasks about memory capacity. If an edit would push a memory over the cap, the write is rejected rather than silently truncated, then Swan makes room and retries.Versioning and restore
Every update to a memory is versioned, with a short summary of what changed and who (or what) changed it. If an edit goes wrong or removes something you wanted to keep, a prior version can be restored.Memory vs. org knowledge vs. skills
- Org knowledge is structured (value prop, competition, ICP segments) and drives how Swan researches and writes outreach at a category level.
- Memory is unstructured — a running set of notes at the org, user, or record level.
- Skills are instructions for how to do something, not facts to remember. See skills.
Common questions
Who can see a record’s memory? Anyone in your org working with Swan on that record can read it, the same way any other Swan data is scoped to your org. In the brain, record folders need Record read permission to open and Record update permission to edit. Can I edit a memory directly instead of asking Swan to? Yes — you can open a memory and edit the text yourself at agent.getswan.com/memory (or, for a company, its memory card/page; for other records, ask Swan); Swan also writes to it conversationally as it learns things. What happens if my memory is already full and I need to add something important? Ask Swan to consolidate it first — it will tighten existing notes to free up space — or manually trim less important content before adding the new note. Does org memory apply to every user in the org, or just Admins? Every user’s conversations load org memory as context; only Admins can write to it. Is there a memory for individual contacts, separate from companies? Yes. Every record, including each contact and deal, has its ownMEMORY.md in its record folder, separate from its company’s.
Can I see the version history of a memory?
Yes, past versions are kept and can be restored if needed.