Every AI you use, finally knowing who you are.
One approved memory layer above the models—grounded in the work you already do, portable to any compatible client, and owned by you. Correct it once and every connected agent inherits the correction.
If the context lives inside one model, it is not shared memory.
Every assistant builds a different picture of your team. Switch tools and the context stays behind; add an agent and it starts from zero. Jackdaw keeps the durable memory above the models, where your team can inspect, correct, and own it.
Useful, but siloed.
Each assistant learns a different version from activity inside that provider. Your team cannot inspect or share one canonical memory.
Portable, but manual.
You decide exactly what goes in, then keep it current and copy the same context into every tool yourself.
Shared, visible, and yours.
One editable wiki, typed and sourced facts, reviewable proposals, and the same approved memory delivered to compatible clients over MCP.
One durable memory layer under every connected AI.
Fresh evidence flows in. Approved knowledge persists. Relevant context flows out wherever your team works.
Connected work keeps it grounded.
Supported email, notes, calendar, and code connections bring fresh evidence into Jackdaw in the background. Conversation captures and external-AI proposals remain reviewable.
Native connectors for common sources. Anything with an MCP server works too.
You decide what becomes permanent.
When an agent learns a decision, constraint, or preference, it can propose the update to your Inbox. Approve it once and future sessions can inherit it.
“API layer uses tRPC exclusively. No REST endpoints.”
Every connected AI starts from the same approved context.
Connect Jackdaw over MCP in about two minutes. Claude, ChatGPT, Gemini, Cursor, and any compatible client read the same current context.
$ claude mcp add --transport http jackdaw https://mcp.jackdaw.so
✓ jackdaw · http · connected
# every session now starts with your context
Pull it. Push it. Keep the session on course.
Jackdaw does not only store memory. It delivers the right part of that memory into the AI session—and can stay present while the work unfolds.
Let the agent ask for context.
Connect any compatible AI client over MCP. It can retrieve the relevant slice for the question, search the wiki, and propose what it learns back to your Inbox.
Best for Claude, Cursor, custom agents, and any client that speaks MCP.
Connect over MCPPut context into every prompt.
The Claude Code hook injects relevant projects, decisions, constraints, and recent activity before each prompt—even when the model would not have thought to ask.
Fails open if Jackdaw is unavailable, so it never blocks your work.
Install the Claude Code hookKeep a long-running session on course.
A Jackdaw agent can observe Claude Code events with your context in view. At a stop, it can let the session finish or return a continuation instruction tied to the goal you gave it.
You define the guardrail—for example, “do not stop until the tests pass.”
Build a supervisorCorrect it once. Every connected AI can start from the approved truth.
A correction stays in the current conversation first. Approve it for your long-term wiki and compatible clients can retrieve the corrected fact instead of the value it replaced.
I left Acme. I’m at Northstar now.
Current employer: Northstar
Acme superseded · history retained
Connected over MCP · model responses can still vary
Already have an agent? Give it your memory.
Jackdaw speaks MCP, so anything that does can read the same approved memory—Claude, ChatGPT, Cursor, Copilot, or the one you wrote yourself. You are not moving in. You are giving what you already use a memory it never had.
Claudereads your decisions before it answersChatGPTdrafts in your voice, not a generic oneCursoruses the retention you told the auditor
Underneath it all, a purpose-built engine keeps that memory useful as it grows.
Not more context.
The right context.
The obvious move is to throw everything in: every rules file, note, and decision. Past a certain scale, models stop prioritizing; an obsolete rule can outrank the one you meant, and the answer comes back confidently wrong. Jackdaw sends only what the question needs.
“I'll wire up a REST endpoint — your /services/api/CLAUDE.md says REST is fine for public routes…”
“Use tRPC. Your canonical rule is tRPC-only — the ‘REST ok for public’ note was retired in Q3, so it doesn't apply.”
Side-by-side example — illustrative
We test what happens when the haystack grows.
We buried answers to 46 realistic questions in a synthetic corpus that grew from 2,000 to 128,000 distractor records, then compared the same retrieval approaches at every scale.
answer correctness · internal synthetic benchmark · one run · Jun 2026
Jackdaw’s resolved memory degraded less in this run. It stayed between 86% and 94%; vector and hybrid RAG fell from roughly 80% to roughly 60%. The shape is promising, not a guarantee—and we keep the caveat attached.
Memory has to know what is current.
Simple retrieval can surface every version of a fact and leave the model to choose among contradictions. Jackdaw resolves revisions into a canonical current value while preserving the evidence and history behind it.
Both facts are retrieved. The model has to guess which one still applies.
The current fact is explicit. The old value remains traceable instead of competing for the answer.
See the rest of Jackdaw.
Why a jackdaw? One of the smartest birds alive — it seeks out what’s valuable and keeps it close. Jackdaw follows the same principle: search a large connected history for the few sourced facts most relevant right now. The right context, not all of it.
Give your agents a memory worth keeping.
Jackdaw is in closed beta. Request access and we’ll reach out personally when a spot opens.