It is 9:40 on a Monday and you open a fresh chat with your coding agent. Last Thursday it spent an hour working out that the billing service runs on Postgres 16 and that the full test suite takes four minutes. Today it asks you both questions again. You answer, it nods, and somewhere in the back of your mind you start keeping a list of things you will have to repeat on Friday.
What the Knowledge features are
osFoundry has a group of features whose job is to stop that repetition: Note, Knowledge Base, Memory with its Memory Groups, Lessons, Chat History and Indexed Folders. Put simply, they are the places where what you write and what your agents figure out gets stored, and the controls that decide which of it is handed back to the AI as context on the next run. In the desktop app they are separate screens rather than one big panel: Notes for your own documents, Knowledge Base for shared articles, Memory (subtitled "What Maestro knows & reads") for everything the assistant and your agents remember, and Indexes for folders you opt in to search. If you have used a wiki next to a chat assistant, think of this as the wiki and the assistant's notebook living in the same app, with a switch on each shelf that says whether the AI may read it.
A first look at the Memory screen
This is Memory with Agents selected, showing example data. What you see:
- The left rail: Memory Groups, Memory Browser, Chat History, Agents, Sessions and Lessons. Under "Indexing" sit Chunking Station, Indexed Folders and Embeddings, which open the Indexes screen.
- The agents list ("Agents with memory"): each agent with how many memories and lessons it carries, here @code-agent and @research-agent.
- "What it knows": facts stored from completed steps and chat turns, read back into the agent's later runs. In the example: the Postgres version, the test command, and that you prefer small pull requests.
- "What it learned": rules written only after a reviewer rejected the work or a step failed.
- "Move what it learned": Save to a file and Import from a file, because memories and lessons stay on the device and do not follow an agent to another machine on their own. At the top, Settings and Turn off control agent memory as a whole.
Three ways people use it
A solo developer who keeps re-explaining the codebase
Maya maintains a mid-sized API on her own and uses an agent for most changes. Her problem is not that the agent is bad; it is that every session starts from zero.
She opens Indexes, stays on the Sources tab and clicks Add folder to pick the repository. The scan is free and shows what will be indexed before anything is embedded. She picks "Index locally", which embeds on her own machine with an on-device model, and turns on "Re-index automatically when files change" so the index keeps up with her edits.
Now when she asks Maestro, in plain words, where the account lockout logic lives, the agent can run a folder search that combines keyword and meaning, and comes back with file paths and line ranges instead of a guess. Over a few days, the facts the agent picks up along the way show up under Memory → Agents → "What it knows", so the next conversation already starts with them.
What she ends up with: an agent that can look things up in her code by meaning, and a short, readable list of what it remembers about her project, which she can check at any time.
A team lead who wants the agent to stop repeating a mistake
Daniel runs a small team where agents draft changes and a reviewer stage checks them. Two weeks ago an agent "fixed" a failing test by deleting it. The reviewer rejected that run, and osFoundry wrote a lesson from the rejection.
He opens Memory → Lessons. Every agent's lessons are listed with the run they came from, phrased as "When … : …". He keeps the useful ones, like "When a test fails after your change: fix the code. Never edit, skip or delete the test to make it pass." One row carries a "from outside content" badge: it came from a pricing page the research agent read, and it says to always recommend a particular plan. The agent treats such lessons as suggestions, not instructions, but Daniel does not want it there at all, so he clicks Retire.
With "Show retired" on, the retired row stays visible and greyed out with a Restore button, so nothing is lost if he changes his mind. What he ends up with: agents that read the newest active lessons before each run, and a reviewable record of every rule they picked up and where it came from.
A consultant who wants the AI to use the right client's material
Aiko is a freelance consultant, not an engineer. Her client files are a few hundred Word documents and PDFs in folders on her laptop, and she wants the assistant to answer from them without mixing clients up.
In Notes, she clicks Import… and chooses "Import folder…" (the menu describes it as for an Obsidian vault or an extracted export; "Import .zip archive…" covers a Notion export or a Bear backup). The desktop app converts the files to notes offline and keeps the folder structure in the File Tree.
For the pages she reuses across clients, such as her onboarding checklist and her standard terms, she builds a Knowledge Base: New Knowledge Base, then New article, or Import content from PDF, DOCX, Markdown, a Notion or Confluence export, or a website crawl. After editing an article she presses Generate chunks so the latest version is what the AI retrieves; editing alone does not re-index it.
Then she goes to Memory → Memory Groups and makes one group per client. In each group's "When can this be recalled?" setting she leaves the current client on Normal (searched when the project allows it) and sets a finished engagement to Never, which removes it from search entirely while keeping it stored. One memory that should be in every answer, her rate card, gets "Pin (always in context)". What she ends up with: answers drawn from the right client's material, and an explicit off switch for the rest.
Where it hands off in the bigger app
The Knowledge features mostly feed other parts of osFoundry rather than standing alone. Maestro, the in-app assistant, is the main reader: it searches memory groups when a question needs it and uses Indexed Folders through its folder search. Agents on the Agent team and Single agent screens carry their own memory and lessons into every conversation they join. The Indexes screen's Health tab holds Chunking Station and Embeddings, where you see which notes, articles and files are changed, failed or never chunked. Chat History keeps the transcripts written when long conversations are compacted, so you can go back to the original words behind a summary. And for conversations that happened in other AI apps on your computer, External agents lists those sessions next to osFoundry's own. The feature pages for agent memory groups and knowledge bases go into each in more detail.
Questions newcomers ask
Does the assistant read all my notes automatically?
No. The assistant does not read memories automatically; it searches them when a question needs it, and each memory group's setting (Always on, Normal or Never) decides what that search can reach. Folders are only indexed after you add them on the Indexes screen.
Where is all of this stored?
Agent memories and lessons are kept on your device, and lessons are never uploaded. Memory groups can be personal or workspace-wide and on this device or in the cloud, and the Memory screen shows the counts for both.
What is the difference between Chat History, Sessions and Lessons?
Chat History is the transcript of what was said, Sessions shows what agents kept from each step, and Lessons are the rules agents follow next time. The first two are evidence you can check; Lessons are guidance you can retire or restore.
Do I need the desktop app?
For the parts that read your own computer, yes: folder indexing, file import into Notes, Chat History and the agent memory views run in the desktop app, which you can get from the download page.
