How Jackdaw works
The model behind your memory — subjects, typed facts, provenance, and how corrections supersede over time.
Jackdaw turns the things you tell it and the data you connect into a structured, lasting memory that any AI can read. Six ideas explain the whole system.
Your wiki is made of subjects
Everything Jackdaw knows lives in your wiki — a set of linked pages called subjects. There's a page for you, and a page for every person, project, organization, decision, and goal that matters. Each page mixes prose (readable, editable) with structured facts, and links to related subjects so your world is a graph, not a flat list. Using the wiki covers the surface; this page is the model underneath.
Facts are typed
Under the prose, Jackdaw stores discrete facts — and each one has a type, because the type changes how an AI should treat it:
| Type | Means | Example |
|---|---|---|
fact | something true about you | "Works at Stripe" |
preference | what you usually want | "Prefers TypeScript" |
constraint | a hard rule to always follow | "Never use any in this project" |
decision | a choice you made | "Chose Postgres over MySQL" |
goal | something you're working toward | "Ship v1 by Q3" |
rejection | something you don't want | "Don't suggest Slack" |
correction | a fix to something Jackdaw believed | "Not the founder — the CTO" |
A preference an agent may override; a constraint it should not. Typing the fact is what lets an AI tell the difference instead of guessing.
Every fact traces back to a source
Jackdaw doesn't store free-floating claims. Each fact carries its provenance — where it came from. On a subject page, Linked content shows the emails, notes, messages, and events that mention that subject, so you can always see why Jackdaw believes something. Facts you state directly are sourced to that statement. It's the "citations or it didn't happen" rule, applied to your own memory.
Facts change over time — corrections supersede, they don't erase
This is the part that makes Jackdaw different from a pile of notes. A fact has a lifespan: a time it became true and, if it's later replaced, a time it stopped being true.
When you correct something, Jackdaw doesn't overwrite the old fact — it supersedes it. The new fact becomes current; the old one is kept, marked with the date it stopped applying.
"I left Datadog — I'm at Stripe now."
Ask "where do I work?" and you get Stripe. Ask about your history and Datadog is still there, correctly marked as past. This is why an AI grounded in Jackdaw doesn't drift back to stale facts — and why a vague correction ("that's not right") doesn't help: Jackdaw needs the new value to supersede the old one. See Correct early, correct explicitly.
Nothing lands without your say-so
Jackdaw learns constantly — from your data, your conversations, and connected AIs — but it never writes to your profile silently. New facts, corrections, and conflicts wait in your inbox for a one-tap Approve or Reject. (Facts you state directly to the assistant are saved right away; the inbox is for corrections and for anything a connected tool proposes.) See Reviewing what Jackdaw captures.
The result is a memory that's curated, not a firehose — and one you own.
How an AI actually reads it
When a connected AI asks Jackdaw a question, it doesn't get your whole profile. Jackdaw works out what the question is about and returns a ranked, relevant slice — the people, projects, and facts that matter for this question, within a token budget. Reads are read-only; the only way an AI changes your memory is by proposing a fact, which lands in your inbox. Connect via MCP covers the tools.
It's all scoped to a workspace
Every piece of this lives inside a workspace — your personal one, a family one, or a team one. Facts, sources, and subjects never cross between them. See Workspaces & teams.