The collaborative brain for humans and agents.

Your team's shared memory: cited, current, and compounding. Ask it anything. Every answer arrives with receipts.

⌘K

Grounded in

●●●●●

Your context is everywhere.
Which means it's nowhere.

Your company already wrote down what it knows. Meanwhile every agent you hire starts from zero, asks the same questions, and forgets the answers by the next run. The work compounds. The context doesn't.

cost of no memory: one question, one quarter

"where did we land on the EU rollout?" #general · mar 4

asked again in standup meet · mar 11

asked again by the new hire DM · mar 26

asked by the support agent, answered wrong 40 min lost · apr 2

with Sift: asked once, answered with receipts S2 S7 · 0.4s

One brain. Every source. Receipts included.

Connectors feed meetings, documents, code, and conversations into one brain. Every record keeps its source, its timestamp, and its history. Watch it keep itself current.

meeting ended · 3 records updated

Launch readiness, payments migration decision

Cutover is go for June 17, pending the ledger backfill.S2

Backfill dry run finished clean across all 4 shards.S4

Rollback window is 20 minutes.S5

Support macros drafted by the docs agent, awaiting review.S3

src:slack/#eng-payments · 2m ago · grounded in S2 S4 S5 ●●●●○

Your agents wake up knowing everything.

People use the app. Agents come in through MCP, CLI, and API: same memory, same citations, same permissions. What one learns, all of them know.

MCP CLI API

agent session · authenticated

Yes. Cutover is go for June 17, pending ledger backfill [S2]

Heads up: rollback window is 45 min, corrected by Maya [S5]

via MCP · 240ms · grounded in S2 S5 · skill:verify-rollback-windows@v2

Already running a coding agent? Hand it the brain.

Paste this into Claude Code, Cursor, or any agent with a shell. It pulls the Sift skill, sets up the CLI, and registers itself, so it wakes up knowing what your company knows.

paste into your agent

You now have access to Sift, our shared cited brain. To set up, run:

  npx -y @sift-wiki/cli@latest skill install

then open the sift-setup skill it installs and follow it. It signs you in,
registers you as an agent on the workspace, and connects you to the brain.

No Node yet? Get the skill directly:
  curl -fsSL https://sift.wiki/agent/sift-setup.SKILL.md

Same memory, citations, and permissions as your team. One skill, and every agent you run knows the company.

“Why not just wire every connector into the agent?”

Fair question. Here's the same agent answering the same questions two ways, at the same moment.

⌘K

any agent · 14 connectors0.0s

same agent · one connector: sift0.0s

Fifty connectors give an agent access. Not memory.

Recall stops at each vendor's search box.

A connector can only surface what Slack or Drive's keyword search returns. Phrase the question differently and the answer never comes back. Sift retrieves over one index built for meaning, not keywords.

It forgets everything between sessions.

Same searches, same dead ends, same mistakes, every single run. Sift's corrections persist and version, so every answer after Maya's fix carries it.

One token sees everything.

An agent wired to org-wide connectors shows whatever its credentials can read to whoever asks. Sift filters every answer by the asker's own permissions, at the retrieval layer.

You pay to re-read the company, every question.

Query-time fan-out re-reads raw documents through a frontier model on each ask. Sift understands once at ingestion; questions hit an index, not ten rate-limited APIs.

Keep your agents. They're great hands. Sift is the memory — one connector instead of fifty.

Your experts, distilled into skills that outgrow them.

Sift turns what your best people know into versioned skills, written by experts or mined from your team's own history. Agents exercise them on real work. Experts correct them where they already talk. Outcomes grade them from your systems of record. Every cycle, the skill climbs a version and needs fewer corrections: a closed loop that keeps running until the skill is better than any one expert who taught it.

skill:refund-policy v1

corrections per 100 runs 14

skill created · authored by maya

skill:refund-policy v1

Refunds: 30 days on all plans. Route requests through billing → refunds.

agent run · support · skill@v1

"what's our refund policy?"

Refunds are 30 days on all plans. skill@v1

nobody knows it's stale yet.

correction · slack/#support · 2 days later

maya: "actually it's 60 days on Pro and Scale now"

digest for the owner

- refunds: 30 days on all plans + refunds: 60 days on Pro and Scale

replay: 3 of 12 past answers change · maya ratifies

agent run · support · skill@v2

"what's our refund policy?"

Refunds are 60 days on Pro and Scale. skill@v2

correction · doc edit · sam

"Starter plans stay at 14 days. The old doc said 30."

digest for the owner

- starter plans: 30 days + starter plans: 14 days

replay: 1 of 9 past answers change · sam ratifies

three weeks in · skill@v3

412 runs · 0 corrections

the loop never stops. it just gets quieter.

outcome watcher · src:zendesk

refund escalations ↓ 38%

Better than any one expert who taught it.

verified server-side, never agent-reported

Nothing is taken on faith.

Every claim arrives with its receipts: source, time, and confidence. This is one answer, taken apart.

The migration shipped two days earlyS1 after the team agreed to defer the EU rolloutS2, a call the on-call agent now cites every time someone asks about timelines.S3

src:linear/SIFT-204 2026-05-28 · shipped ●●●●○

src:meet/launch-sync 2026-05-26 · decision ●●●○○

src:agent/run#4821 2026-06-09 · cited 14× ●●●●●

src:notion/support-runbook may 12 · "rollback window: 20 minutes"

contradiction · surfaced, not averaged

src:maya · correction jun 3 · "45 minutes" · human-verified

When sources disagree, the answer shows both and why the newer, human-verified one wins. When it doesn't know, it says so.

Frequently asked questions

Search tools retrieve; Sift remembers and learns. Every answer carries its sources, timestamps, and confidence. When your experts correct one, the correction becomes a versioned skill that every future run inherits. Retrieval is where the others stop; the loop is where Sift starts. And agents are first-class: the same brain your team uses is the one your agents call through MCP, CLI, and API.

Every claim is grounded in a source you can open. Provenance, time, and confidence are first-class, not an afterthought. When sources disagree, Sift surfaces the contradiction instead of papering over it. And when it doesn't know, it says so and shows what it would need to find out.

Visibility follows the source. Knowledge mined from a private channel stays as private as that channel, and an answer never cites something the asker couldn't already read. Agents get the same permission boundaries as the people they work for. Your data is never used to train models.

You correct it where you already work: a Slack reply, a document edit, a teach action. Corrections are batched into a digest with replay evidence showing exactly which past answers would change, the owning expert approves or rejects, and every future run inherits the fix. Every version is reversible in one operation.

No. Sift sits across the tools you already use and reads from them. Nothing has to move, and nobody has to change where they write. The brain itself is markdown-canonical, so what Sift builds stays portable. Your knowledge is yours, not locked in.

Almost nothing. Any agent that speaks MCP (Claude Code, Cursor, your custom stack) gets the brain as a tool: same memory, same citations, same permissions as your team. There's a CLI and API for everything else. One integration, and every agent you run wakes up knowing what your company knows.

Days, not quarters. Connect a few sources and the brain assembles itself from your team's history, including draft skills mined from how your team already works, ready for an expert to adopt. Cold start is never a blank page.

The AI-native operating system companies run on.

Memory is the substrate. Learning is the loop. Agents are the surface. Teams that compound their context will outrun teams that don't. We're building the brain they'll run on.

Book a demo charles@sift.wiki

A brain built from layered paper, with amber threads connecting points across its regions

Your context is everywhere.Which means it's nowhere.