01 / CAPABILITY
SE
Session-aware updates
Give each AI chat or task a readable session name, then associate its memory entries with that session.
MCP memory server / Hosted / Cross-client
Hosted MCP memory server.
Shared by Cursor, Claude, and Codex.
Create one searchable, append-only project memory and connect compatible AI clients through a hosted HTTPS MCP link. Free includes 1 Agent Memory connection and 1 MCP link.
Free: 1 memory + 60 appends/hourProject-scoped MCP linksRules and summaries with review
SHARED MEMORY / LIVE3 AI CHATS / 1 HOSTED MCP LINK
AI CHATS
Claude search
Cursor append
Codex handoff
HOSTED MCPdatamcpauth · access · tools
SHARED PROJECT RECORDAgent Memorystructured · searchable · append-only
Claude
search_memory8 entries foundCursor
append_memoryprogress savedCodex
create_handoff_contexthandoff readyACCESSPer MCP link
MODERead / Append
STORAGEHosted
FORMATStructured
See the handoff
One fact. Three chats.
No copy-paste.
Agent Memory is a shared project record. One AI client writes a useful update; another can find it later and continue with the decision intact.
ClaudeAI CHAT
We will ship the append-only memory model.
DATAMCP TOOL
→append_memorySHARED MEMORYDecision saved with the project, files, and session.
Free: one Agent Memory and one hosted MCP link.
Browsing on your phone? Create the account now and connect your desktop AI client when you are back at your computer.
Direct answer
What does a hosted MCP memory server do?
A hosted MCP memory server gives compatible AI clients a shared, tool-accessible project record without requiring you to operate a local database, custom MCP runtime, or synchronization service.
- Best fit
- Projects where multiple AI chats or agents need to leave durable progress, decisions, blockers, deploy notes, and handoff context for one another.
- Cross-model continuity
- Start in one compatible AI client and continue in another using deliberately saved project memory instead of rebuilding the context from a transcript.
- Project model
- One memory source can contain separate projects, and each MCP link can be scoped to one project.
- Available tools
- Project, session, append, canonical context, proposal, search, entry retrieval, handoff, and compaction tools. Visibility depends on link access.
- Storage model
- Append-only work logs plus curated Rules and Project Summary documents. Compaction can archive source entries from default results without deleting them.
- Search model
- PostgreSQL full-text search with filters. This release does not provide vector embeddings or semantic recall.
- Write control
- Read Only blocks writes. Read & Append adds logs, compaction, and canonical proposals. Full Access can apply canonical changes for an owner/admin API key user.
- Free limit
- 1 source connection, 1 MCP link, and up to 20 appends per minute or 60 per hour per user and memory connection.
- Not included
- No automatic task-completion writes, verified agent identity, direct .md upload, edit/delete tools for work-log entries, or replacement for Git and project documentation.
The production path
A shared project record that AI clients can use
AI chats often lose the decisions that matter between sessions. One chat changes an API contract, another investigates a production bug, and a third starts without knowing either happened. Copying entire transcripts is noisy and usually includes more context than the next task needs.
Agent Memory gives each compatible client explicit Model Context Protocol tools for projects, stable context, logs, search, and handoffs. A client reads the project's curated Rules and Project Summary, registers a human-readable session name, appends meaningful updates, searches earlier entries, compacts older work, and requests focused handoff context when another chat needs to continue.
Your project memory does not have to stay locked inside one AI subscription or native chat history. Work can begin in Claude, continue in Codex, and move to Cursor when each client is authorized to use the same MCP link. The next client receives the structured facts and handoff that earlier clients deliberately saved, not private reasoning or an automatic copy of the conversation.
The current system is deliberately structured and append-only. It records what an agent chose to write, not every conversation automatically. Search uses PostgreSQL full-text search and filters rather than embeddings. That makes the behavior understandable, but it also means clients need clear instructions about when and what to save.
Read what agent memory is and when to use it, or follow the Agent Memory setup reference for tools, permissions, and limits.
What you get
Production access.
Without production guesswork.
02 / CAPABILITY
SE
Separate project namespaces
Keep products, repositories, clients, or workstreams apart inside one Agent Memory source and scope an MCP link to one project.
03 / CAPABILITY
CA
Canonical Rules and Summary
Give each project stable context that trusted clients can read, propose changes to, and review before direct application.
04 / CAPABILITY
ST
Structured project memory
Store summaries, Markdown details, decisions, risks, deploys, tags, files, importance, and metadata in a consistent entry model.
05 / CAPABILITY
FI
Filtered full-text search
Search project history by text and narrow results by date, tags, files, type, importance, or session.
06 / CAPABILITY
CO
Compact handoff context
Create a bounded summary of recent or filtered entries instead of pasting full chat transcripts into the next agent.
07 / CAPABILITY
CR
Cross-model continuity
Continue a project across Claude, Codex, Cursor, or another compatible MCP client without making one vendor's native chat history the only copy of project context.
08 / CAPABILITY
DU
Durable memory compaction
Turn selected active entries into a structured summary and optionally archive the source entries from default results without deleting them.
09 / CAPABILITY
PE
Per-link write control
Use Read Only for retrieval or Read & Append when a trusted client should record new project context.
010 / CAPABILITY
HO
Hosted remote MCP
Connect through one HTTPS MCP endpoint without running a local memory server or maintaining its storage layer.
Source to controlled MCP link
One. Two.
Three.
Create the memory once. datamcp hosts its storage, remote MCP transport, client authentication, search tools, and per-link write boundary.
Read the setup guide- 01Create Agent Memory
Add an Agent Memory source, then use its default project or create separate projects for distinct workstreams.
- 02Choose project and access
Scope the link to one project when needed. Use Read Only for recall, Read & Append for coordination, or Full Access only for trusted canonical review.
- 03Connect each AI client
Add the hosted MCP URL to compatible clients and instruct each chat to read Rules and Project Summary before registering its session.
Build or use datamcp
Hosted Agent Memory vs. copying chat context
See what stays on your engineering backlog when the MCP runtime and control layer are managed for you.
CAPABILITYdatamcpManual notes and transcripts
Shared accessOne MCP memory for compatible clients
Copy and paste between chats
Context shapeStructured summaries and Markdown entries
Long transcripts or inconsistent notes
Stable contextCurated Rules and Project Summary
Mixed into temporary notes
RecallFull-text search with filters
Manual document or chat search
HandoffsSummaries plus recent filtered entries
Rebuild context by hand
AI client changesReuse saved context across compatible MCP clients
Restart from a transcript or manual brief
Write boundaryRead Only or Read & Append per link
Controlled by access to the whole document
Frequently asked questions
Questions before
you connect.
Clear answers about architecture, access, credentials, and supported clients.
What is an MCP memory server?
An MCP memory server exposes memory tools through the Model Context Protocol. AI clients can register a session, append structured project updates, search previous entries, retrieve a full entry, prepare handoff context, and compact older entries into a durable summary.
Is datamcp a local MCP memory server?
No. datamcp Agent Memory is a hosted remote MCP server reached through an HTTPS MCP link. Use a local memory server instead when the storage and MCP process must remain entirely on one machine or private network.
How does Agent Memory keep a long project history compact?
The summarize_memory tool creates a structured Markdown summary from selected active entries. Source entries can stay active or be archived from default search and handoff results. Archived entries are not deleted and can still be retrieved explicitly.
What can an AI agent store in datamcp Agent Memory?
An entry can include a title, summary, Markdown body, entry type, importance, tags, file paths, metadata, session ID, and occurrence date. Entry types include notes, progress, decisions, bugs, handoffs, deploys, tasks, and risks.
Is this semantic or vector memory?
No. The current search uses PostgreSQL full-text search plus filters such as tags, files, entry type, importance, session, and date range. Do not treat it as embedding-based semantic recall.
Can agents edit or delete work-log entries?
No. Work-log entries are append-only. Agents can append, search, retrieve, compact, and hand off those entries, but cannot update or delete them. Curated Rules and Project Summary use a separate proposal and review workflow.
Does datamcp automatically write after every task?
No. The connected AI client must call append_memory as part of its own instructions or workflow. datamcp provides the tools and hosted storage, but it does not force a client to record an update.
Can I upload a Markdown file and turn it into an MCP server?
Not in the current release. Memory entry bodies support Markdown, but direct .md file import, synchronization, and export are not available.
How are memory writes controlled?
Use Read Only for project listing, canonical reads, search, retrieval, and handoffs without writes. Read & Append adds sessions, logs, compaction, and canonical change proposals. Full Access can apply canonical changes directly for an owner/admin API key user. Each MCP link can be scoped to one project and revoked individually.
What are the Agent Memory limits?
Per user and memory connection, Free supports 20 appends per minute and 60 per hour, Pro supports 120 per minute and 2,000 per hour, and Enterprise supports 600 per minute and 20,000 per hour. Read, handoff, compaction, project, session, and canonical-update rates also scale by plan. Additional connection-level and organization-level safety limits apply.
Does Agent Memory replace Git, documentation, or issue tracking?
No. It is a compact coordination layer for AI chats and agents. Keep source code in version control, durable documentation in its canonical system, and formal work items in the appropriate tracker.
Can I continue the same project in another AI client?
Yes, when both clients support the required remote MCP connection and are authorized to use the same Agent Memory link. The next client can read deliberately saved project entries and handoff context. This does not transfer hidden chat history, model state, unsaved reasoning, or account access, and it does not bypass any provider's usage limits.
Keep exploring
Related guides.
GUIDE
Codex memory and MCP
Compare Codex Memories, AGENTS.md, task history, and a shared MCP memory server.
GUIDECursor memory and MCP
Compare Cursor Memories, Rules, file-based Memory Banks, and hosted shared memory.
GUIDEMulti-agent memory architecture
Design shared context, write boundaries, consistency, and handoffs across several AI agents.
GUIDEClaude persistent memory options
Compare Claude Memory, project memory, CLAUDE.md, Claude Code auto memory, and a shared MCP memory server.
GUIDEMemory Bank MCP comparison
Compare file-based memory banks with a hosted, structured, append-only Agent Memory service.
GUIDEWhat is agent memory?
Understand persistent context, shared project memory, MCP tools, and the limits of the current approach.
GUIDEAgent Memory documentation
Configure projects, canonical context, MCP-link access, tools, content limits, and plan rate limits.
GUIDEHosted MCP gateway
See how datamcp hosts remote MCP endpoints and controls client access across supported source types.
GUIDEMCP permissions and scopes
Separate client identity, MCP-link permissions, backend access, and revocation.
Start with one link
Keep project memory independent of one AI client
Create Agent Memory in datamcp, receive a hosted HTTPS MCP link, and let authorized compatible AI clients continue from the same structured project record. The Free plan includes 1 connection and 1 MCP link, with no credit card required.