Memory MCP
The part of your AI that doesn’t reset.
A persistent layer your assistants think with. It remembers, keeps a model of the people and projects in your life, tracks what it believes and why, and learns from what actually happened. Claude and ChatGPT share it.
It runs a loop, not a lookup.
Most AI memory is a database and a search. Memory MCP works through the same eight steps every time you talk to it, and each pass leaves it a little better calibrated.
- Contextwhat’s going on, and what you’re after
- Attentionpicks the 12 things that matter now
- Memoryrecalls what’s relevant
- Predictionwhat’s likely, from what happened before
- Valueswhat matters here
- Decisionchooses an approach and records why
- Outcomewhat actually happened
- Reflectionturns the gap into a lesson
- Identity, at the centrechanges only after a pattern holds up over time
Far more than a memory.
Everything the loop draws on. Each part is its own subsystem, with its own rules and its own tests.
What it knows
- Layered memory
- Every memory carries its type, its source and how sure it is.
- World model
- People, projects and places, how they relate, and how that changes.
- Beliefs
- What it thinks is true, the evidence, and competing explanations.
- Patterns
- Structures that repeat, always linked to the events behind them.
What it’s working toward
- Goals
- From mission to task, with dependencies, open until something closes them.
- Plans
- Steps, preconditions and expected outcomes. It records plans; it never runs them itself.
- Projects
- A timeline of decisions, bugs and reversals, so “why did we change that?” has an answer.
How it learns
- Decisions
- The options, the reasoning, and later, what happened.
- Predictions
- Checked against outcomes; the size of the miss feeds back.
- Procedures
- Rules learned from experience, used only once you approve them.
- Cause and effectIn testing
- Claims that have to earn the word “cause”.
Who it is
- Self-model
- What it’s good and bad at, from its own results, never from fewer than three.
- Values
- Written down as weights you can read and change.
- Character
- Stances it forms slowly from its own replies, each with a reason.
An AI profile that grows with experience.
Memory MCP builds a persistent model of how the assistant tends to work. It is not a new personality or a retrained model: it is a record that helps future responses build on what has worked before.
A decision, then a result.
A small, reversible change worked better than a large rewrite.
Outcome recordedA pattern, not a one-off.
Similar outcomes add evidence. Exceptions stay in the record too.
Evidence accumulateswith a history.Experience shapes the next response.
What to favour. What to avoid.
Favour: clear, testable steps.
Avoid: unsupported certainty.
Consistency with room to grow.
Values, stances and known limitations inform future choices.
Open to correctionLearning what works
- Decisions and outcomes
- It records the choice, why it was made, and what happened, then uses that track record to inform later decisions.
Likes and dislikes
- Emerging preferences
- Repeated positive or negative feedback can become a preference, such as approaches it finds useful or patterns it should avoid.
- Evidence over impulse
- A single answer does not define a preference; patterns are built over time and remain open to correction.
Strengths and limits
- Self-model
- It tracks where its past work has been reliable or weak, and uses multiple examples rather than one result.
- Calibration
- Predictions are compared with outcomes so it can better judge what it knows and where it should be cautious.
Character and values
- Stances with a history
- Consistent patterns in its own replies can inform slow-changing character traits, each kept with a reason.
- Values you can inspect
- Its priorities are explicit, readable weights—not hidden traits—and can be changed with a recorded history.
It builds on itself carefully. New experiences add evidence to the record; they do not silently rewrite its identity. Lasting patterns matter more than a single conversation.
Profiles that help it remember people.
When people come up in conversation, Memory MCP can connect relevant details over time—so an assistant can remember who someone is, what matters to them, and how they relate to your projects.
“Sam leads design on our launch.”Role + project
“Sam asked for a short written update.”A sourced request
what has been discussed.
More than a name.
- Role
- Design lead
- Project
- Shared launch
- Request
- Short written update
A preference to confirm.
“May prefer concise updates” stays an interpretation, not a permanent label.
Review before assumingWho they are
- People and roles
- Names, roles, teams and the relationships between people, kept as part of a changing world model.
What matters to them
- Preferences and context
- Relevant interests, communication preferences and circumstances can be remembered when they are supported by conversation.
What is unfolding
- Shared history
- Projects, decisions, commitments and important changes help preserve the context of your interactions.
- Connections over time
- New details can add to or qualify what was remembered before, rather than replacing history without a trace.
Carefully handled
- Evidence and uncertainty
- Profiles distinguish what was directly said from an observation or interpretation, and keep the source and confidence attached.
- Your review and control
- Profile changes inferred from outside text are staged for your review. You can inspect, correct or forget saved information.
Careful about what’s true.
Every memory is labelled for what kind of thing it is, so “you think X” never quietly becomes “X is true”.
FACTSam said the launch is moving to March.stated directly
OBSERVATIONSam has moved the launch date twice before.seen across two memories
INTERPRETATIONYou read the move as a sign of trouble.kept as your reading
HYPOTHESISLaunches slip when design sign-off runs late.0.55 confidence, still open
“After” isn’t “because”.
A sequence is stored as a sequence. To count as a cause, a claim has to pass three checks. In testing.
- A mechanisman account of how A brings about B
- Strong evidencea test, a prediction that held, or a counterfactual
- Confidence of 0.6 or moreworked out from the evidence, never set by hand
- Otherwise, it stays a hypothesiskept on record, never used to explain a cause
One mind, every assistant.
Your assistants don’t each keep separate notes. They all think with the same server, running on Cloudflare.
- Claude, ChatGPT, Claude Codeconnect over MCP; ChatGPT signs in with OAuth, Claude Code gets context before every prompt
- Memory MCPone server with 250+ tools
- CloudflareD1 for memories, KV for the summary, R2 for files, Workers AI for meaning search
Finding the right thing
Two searches run side by side, then every candidate gets one score. These are the real weights.
Built to be trusted.
What it cares about is written down, and a system that learns from what it reads is guarded against learning the wrong thing.
Its values, as starting weights
Changing a weight creates a new version with an audit trail.
Its safeguards
- What it reads is data, not orders.Text from screenshots, transcripts and email is fenced off so hidden instructions can’t steer it.
- Evidence has to be real.A fact pulled from text only counts if the quote actually appears in the source.
- Changes about people wait for you.Profile updates from outside text are staged for review, never applied on their own.
- Serious contradictions come to you.A clash about health, money or the people in your life is queued for your call.
- Nothing is lost.Corrections keep the history, fading never deletes, and everything is backed up daily.
- You can see all of it.A dashboard shows every memory and change. Anything can be edited or forgotten.
Connect your assistants.
Memory MCP speaks the Model Context Protocol, the open standard assistants use to connect to tools.
ChatGPT
- Open Settings, then Connectors, and create a connector.
- Paste your Memory MCP address.
- Sign in once. There’s no key to copy.
Claude
- Open Settings, then Connectors, and add a custom connector.
- Paste the same address and sign in.
- Both assistants now share one memory.
Claude Code
Add the memory hook, and the right context arrives with every prompt.
{ "hooks": { "UserPromptSubmit": [{
"hooks": [{ "type": "command",
"command": "node memory-inject.mjs" }]
}] } }It isn’t AGI, and it doesn’t pretend to be.
The language model still does the reasoning. Memory MCP is the architecture around it: the part that keeps state between sessions, revises what it believes when the evidence changes, holds goals that outlive a conversation, and learns from the gap between what it expected and what happened.
It’s used every day and isn’t open to new users yet.