Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@Unified Memory MCP save a memory: API key rotation due Friday"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Local, single-tenant memory and handoffs for MCP clients.
The core uses Python's standard library and SQLite FTS5: no account, hosted service or runtime dependency is needed. Optional embeddings are a separate extra and may download model weights; the default installation uses text search.
Python 3.11 or newer; candidate qualification uses Python 3.12.
python -m venv .venv
# Activate .venv using your platform's standard command.
python -m pip install .
python -m unittest discover -s tests -v
unified-memory-admin --helpThe distribution name is unified-memory-mcp. Version 2.5.1 in this snapshot
is the public-safe candidate, not a claim that a package was uploaded.
Related MCP server: mcp-memory-bridge
Use your MCP client's configuration syntax, with explicit executable and database paths. For a client accepting a mcpServers object, the shape is:
{
"mcpServers": {
"memory": {
"command": "/absolute/path/to/.venv/bin/unified-memory-mcp",
"env": {
"UNIFIED_MEMORY_DB": "/absolute/private/path/memory.db",
"MEMORY_AUTOSYNC": "0"
}
}
}
}On Windows the installed executable is under .venv/Scripts. The same user can point multiple local MCP processes at one SQLite store. Do not share a database between mutually untrusted users or tenants.
Tools include memory_save/search/recall, memory_bootstrap, handoff_save/load, memory_policy and explicit-confirmation pruning. Scopes organize records; they are not user identities or RBAC boundaries.
No local assistant memory is imported automatically.
Explicit import needs MEMORY_AUTOSYNC=1 and CLAUDE_MEM_DIR or CLAUDE_MEM_GLOB pointing to a reviewed directory. Importing private text is your decision.
Personal private-store integration and the workstation-specific sync utility are not included.
The optional HTTP gateway requires --workspace-root explicitly. It offers authenticated file operations as well as memory; use stdio when those capabilities are unnecessary.
HTTP is for a trusted loopback/tunnel deployment, not public Internet exposure. Configure MEMORY_GATEWAY_TOKEN or MEMORY_GATEWAY_TOKEN_FILE; do not commit it.
No real database, token, deployment settings, autostart task or SSH material is shipped or activated.
The history-free snapshot contains no workstation paths, personal assistant memory, private records, database files, tokens, raw logs or private settings.
The source manifest records original/exported hashes and AST transformations. The private live server is not modified by building this snapshot.
Use only a fresh synthetic directory for a first test:
unified-memory-admin restart-drill --work-root ./restart-demo
unified-memory-admin write-deployment-profile --output ./deployment-profile.jsonBackups and restore-copy require explicit paths; restore-copy never overwrites an existing destination. Protect database and backup permissions yourself. The strict deployment preflight includes Windows-specific storage/ACL checks; do not infer Linux ACL qualification from portable stdio tests.
See SECURITY.md. This candidate is released under MIT; third-party dependencies and optional model weights retain their own terms.
Mark a handoff complete once the transferred task is finished.
Inputs: id (required), notes.
Audit recent handoff checkpoints for a project.
Inputs: project (required), limit, status.
Load recent unfinished handoffs for a project, normally from another client.
Inputs: project (required), consumer (required), limit, include_own, +1 more.
Create or update a structured work checkpoint so another Claude/Codex session can continue from the same state. Reuses project+source+session_id.
Inputs: source (required), project (required), summary (required), id, +12 more.
Start or resume work across Codex and Claude. Returns the latest active handoff from another client plus relevant durable memory chunks. Call once at the beginning of substantial work.
Inputs: project (required), consumer (required), query, scope, +3 more.
Delete a memory by id.
Inputs: id (required).
Fetch a single memory by id.
Inputs: id (required).
List most recently updated memories, optionally filtered by scope or tag.
Inputs: tag, limit, scope.
Report content-free scope authorization, quota, retention, and usage state.
Inputs: no named parameters.
Delete expired memories and completed handoffs. Requires confirm=true.
Inputs: confirm (required).
Semantic recall from memory: returns the most RELEVANT chunks by meaning (cosine over embeddings), falling back to keyword search if embeddings are off. Prefer this over memory_search when you want the best context for a topic.
Inputs: query (required), k, scope.
Rebuild chunks + embeddings for all memories. Run once after enabling embeddings (fastembed).
Inputs: no named parameters.
Save a memory (fact, preference, decision, context) to the shared local store that both Codex and Claude read. Auto-chunked and embedded for semantic recall. Use for durable info worth recalling.
Inputs: content (required), tags, scope, source.
Full-text (keyword) search of whole memories. Returns ranked matches.
Inputs: query (required), limit, scope.
Pull Claude's file memories (*.md) into the store. Idempotent and cheap: only files whose content changed are rewritten and re-embedded. Runs automatically on server start.
Inputs: no named parameters.
Update a memory's content, tags, or scope by id (re-chunks on content change).
Inputs: id (required), tags, scope, content.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 16 tools
Most tools have clearly distinct purposes, but memory_search and memory_recall overlap as search mechanisms, and memory_bootstrap overlaps with handoff_load/handoff_list by also returning handoff data. Descriptions help clarify preferred usage, so misselection is unlikely for careful agents.
All tool names use consistent snake_case with a predictable domain prefix: memory_ or handoff_. Actions are generally verbs and the pattern is stable across the set.
16 tools is slightly above the typical 3-15 range, but each tool serves a distinct role in memory management or handoff lifecycle. The set is rich but not bloated.
The surface covers memory CRUD, search, recall, admin tasks, and handoff lifecycle, so core workflows are complete. Minor gaps include no direct handoff_get by id, though handoff_load/list and memory_prune cover most needs.
Private-by-default, local-first memory/context/task orchestrator for MCP apps and agents.
Hosted agent memory: store, search, and recall facts across sessions from any MCP client.
Persistent memory, hybrid search and a goal graph for AI agents, over stdio or remote HTTP.
Persistent, inspectable memory for AI agents with lineage, correction, and a hosted MCP endpoint.
