mem0-oss-mcp
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., "@mem0-oss-mcpAdd a memory that I like dark mode."
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.
mem0-oss-mcp
Small MCP bridge for the self-hosted Mem0 OSS server from mem0ai/mem0/server.
It exposes the Mem0 MCP tool names expected by Codex and forwards them to a self-hosted Mem0 REST API.
add_memory accepts an optional idempotency_key (1–128 visible ASCII
characters) when a Sidecar backend is configured. The bridge forwards it as
the Sidecar Idempotency-Key header. Reusing the key with the same payload
returns the recorded result; a different payload conflicts. Direct Core mode
rejects this option because it cannot provide the durable Sidecar contract.
Added memories use the Sidecar's persisted event ID. Event status remains queryable after a bridge restart. The bridge authenticates the client as usual and reads only that event's bounded status/results within its configured project using its private operator credential; this read omits caller attribution because Sidecar's event endpoint requires an operator principal. Memory operations continue forwarding the original caller attribution.
Publishing the container image
Publishing a stable GitHub Release runs Publish MCP release. After testing
and uploading all client archives, it calls Publish GHCR image to build the
exact tagged commit. ghcr.io/steinx/mem0-oss-mcp receives the Release tag and
full commit SHA as image tags. Drafts, prereleases, main pushes, pull requests,
and manual client archive backfills do not automatically publish images.
Image publications share a queue, and automatic builds update latest only
while their tag is still GitHub's latest stable Release after immutable tags
are pushed, immediately before promotion. A delayed older build
preserves the newer alias.
Client archive failures block image publication; rerun the failed release jobs after resolving the failure. Pin a release tag or digest for deployment; publishing an image does not restart the MCP service.
For release tags containing this workflow update, retry publication without recreating the Release by manually dispatching the workflow from that tag:
gh workflow run publish-ghcr.yml --repo SteinX/mem0-oss-mcp --ref 0.1.6A manual run from a tag preserves latest unless push_latest=true is supplied.
Manual runs from main retain their existing behavior: publish the commit SHA
and update latest. Other branch runs cannot promote latest, even when the
input is set. The publisher verifies its source before logging into GHCR.
Related MCP server: Mem0 MCP Server
Configuration
MEM0_OSS_BASE_URL=http://<mem0-host>:<mem0-port>
MEM0_OSS_API_KEY=m0sk_xxx
# Recommended: route memory operations through the control-plane sidecar.
MEM0_SIDECAR_BASE_URL=http://mem0-platform-sidecar:8765
MEM0_SIDECAR_PROJECT_ID=default
MEM0_SIDECAR_REQUIRED=true
# Required with the sidecar's secure default. This is the private operator
# credential, not a dashboard-created client key.
MEM0_SIDECAR_API_KEY=replace-with-private-operator-key
# Optional; defaults to a per-process ID and a five-minute heartbeat.
# MEM0_SIDECAR_INSTANCE_ID=mem0-oss-mcp-1
# MEM0_SIDECAR_HEARTBEAT_INTERVAL_SECONDS=300
MEM0_OSS_MCP_HOST=0.0.0.0
MEM0_OSS_MCP_PORT=8080
# disabled | static | hybrid | core_api_key
MEM0_OSS_MCP_AUTH_MODE=static
MEM0_OSS_MCP_CLIENT_AUTH_URL=http://mem0:8000/auth/me
MEM0_OSS_MCP_CLIENT_AUTH_TIMEOUT_SECONDS=5
MEM0_OSS_MCP_TOKEN=change-me
MEM0_OSS_DEFAULT_USER_ID=codex
MEM0_OSS_DEFAULT_APP_ID=default
MEM0_OSS_LIST_FETCH_LIMIT=5000
MEM0_OSS_BACKEND_LIST_RETRY_LIMIT=1000
# Optional: only set when backend top_k should differ from MEM0_OSS_LIST_FETCH_LIMIT.
# MEM0_OSS_BACKEND_LIST_FETCH_LIMIT=5000MEM0_OSS_BASE_URL is the base URL of your Mem0 OSS REST server. The port is
not assumed.
Clients configure only the public MCP URL and bearer token. MEM0_SIDECAR_*
variables are private bridge/operator settings and must not be copied into
Codex, OpenCode, Cursor, Claude, Hermes, or other client configuration.
MCP client authentication
Mode | Accepted credentials | Intended use |
| none | loopback-only development |
|
| backward-compatible deployments |
| legacy static token or Core API key | migration only |
| any active Core admin API key | steady state |
static and hybrid require a non-empty MEM0_OSS_MCP_TOKEN.
When neither a mode nor a legacy token is configured, startup fails closed.
disabled must be selected explicitly and requires MEM0_OSS_MCP_HOST to be
a loopback address; there is no non-loopback override.
core_api_key validates the incoming bearer credential through Core
/auth/me; the MCP process receives no static shared secret. Core
unavailability returns a sanitized 503, while an invalid or revoked key
returns 401.
Create one named Core admin API key per client in the dashboard's Client Keys
page. Codex and OpenCode continue using MEM0_OSS_MCP_TOKEN as their local
environment variable name, but its value becomes that client's m0sk_... key.
The endpoint and MCP configuration remain unchanged.
When sidecar routing is enabled with a trusted MEM0_SIDECAR_API_KEY, Requests
attributes new rows as Legacy shared MCP key for legacy_static or by the
Core key label and prefix for core_api_key. Without that private bridge
credential, sidecar rejects caller-supplied attribution instead of trusting
it. Historical rows created before attribution remain
Unknown (pre-attribution); they are never guessed or backfilled.
When MEM0_SIDECAR_BASE_URL is set, all memory operations plus entity
list/delete use the sidecar so its durable project/app index stays current.
MEM0_SIDECAR_PROJECT_ID supplies the project boundary, while each tool call's
concrete app_id remains intact. If the sidecar setting is absent, the bridge
keeps the legacy direct-OSS behavior. A sidecar request failure is returned to
the caller and is never retried as a direct write, avoiding accidental
double-writes.
With sidecar routing enabled, the bridge reports a bounded read/write routing
capability heartbeat at startup, during health checks, and every five minutes.
When MEM0_SIDECAR_REQUIRED=true, startup fails unless the private operator
credential can call the protected heartbeat route. Server-side
AUTO_SAFE consolidation requires a current heartbeat. Set
MEM0_SIDECAR_REQUIRED=true in production so a missing sidecar URL fails at
startup instead of silently selecting legacy direct mode.
get_memories fetches a larger backend candidate window before applying local
app_id and metadata filters. MEM0_OSS_LIST_FETCH_LIMIT is the sidecar target
window and, by default, the largest top_k sent to the backend. Set
MEM0_OSS_BACKEND_LIST_FETCH_LIMIT only when the backend limit must differ. The
default target is 5000. Older Mem0 OSS builds may reject list requests above
1000, so the sidecar retries with
MEM0_OSS_BACKEND_LIST_RETRY_LIMIT and returns degraded_fetch_limit: true.
When the fetched backend window is full, responses include truncated: true and
complete: false; consolidation tools should not treat that listing as complete.
Codex should connect to this bridge, not directly to Mem0 OSS:
[mcp_servers.mem0]
url = "https://<bridge-host>/mcp"
bearer_token_env_var = "MEM0_OSS_MCP_TOKEN"The bridge itself does not terminate TLS. Keep its 0.0.0.0:8080 listener on
a private application network and publish remote MCP only through an HTTPS
reverse proxy or gateway.
Codex plugin
Releases include a complete prebuilt Codex package with the native runtime,
skills, hooks and OSS MCP bridge. Install, update and roll back with the bundled
install.py; connection configuration and credentials stay outside versioned
packages. See release installation and updates.
The source generators below remain available for custom builds.
The full-experience generator supports the current official
integrations/codex-plugin shared-runtime layout and the earlier
integrations/mem0-plugin layout. It pins its upstream submodule to Mem0
v2.2.1. Native hooks and standalone CLI workers load the same private OSS
credentials; project recall retains the existing user/app boundary. The
generated plugin uses global Codex hooks once, and its manifest omits native
hooks to avoid duplicate lifecycle execution. Search and explicit remember
skills use the OSS bridge's scoped tool parameters.
This repository also publishes a Codex plugin marketplace at
.agents/plugins/marketplace.json.
To use the checked-in plugin, provide the bridge URL and bearer token in the Codex process environment, then add the marketplace and install the plugin:
MEM0_OSS_MCP_URL=https://<bridge-host>/mcp
MEM0_OSS_MCP_TOKEN=change-mecodex plugin marketplace add SteinX/mem0-oss-mcp
codex plugin add mem0-oss@mem0-oss-mcpFor Codex Desktop, or for any host, port, domain, or token that should live outside the Codex process environment, generate a local plugin instance instead of editing files in this repository. Pass the bridge endpoint and token at install time; the installer writes the token to a local private dotenv file and the generated MCP config stores only the endpoint, token variable name, and dotenv path. The recommended shell flow reads the token from stdin so it does not appear in process listings.
printf '%s\n' "$MEM0_OSS_MCP_TOKEN" | \
python3 plugins/mem0-oss/scripts/install_codex_plugin.py \
--url https://<bridge-host>/mcp \
--token-stdin \
--installThe installer writes a local marketplace under
~/.mem0-oss-mcp/codex-plugins, patches only that generated copy, and then
installs it through codex plugin add. It never writes token values into
.mcp.json, hook commands, or repository files. By default, token values passed
with --token-stdin or --token are stored in
~/.mem0-oss-mcp/codex-plugins/env/<plugin-name>.env with owner-only
permissions. You can still pass --env-file /path/to/bridge.env to choose the
dotenv location yourself.
For the full official Mem0 Codex plugin experience, including skills and lifecycle hooks, use the official Mem0 repository submodule as the upstream plugin source:
git submodule update --init --depth 1 third_party/mem0
printf '%s\n' "$MEM0_OSS_MCP_TOKEN" | \
python3 plugins/mem0-oss/scripts/install_codex_plugin.py \
--url https://<bridge-host>/mcp \
--token-stdin \
--token-env-var MEM0_OSS_MCP_TOKEN \
--with-hooks \
--install--with-hooks copies third_party/mem0/integrations/mem0-plugin into the
generated local marketplace, adds a small Mem0 OSS compatibility layer, and
merges the official Codex hook entries into ~/.codex/hooks.json. Official
hook and skill files stay in the submodule, not vendored into this repository.
The generated full plugin also replaces the upstream /mem0:dream skill with
an OSS manual-trigger variant that checks listing completeness and requires
confirmation before applying changes.
To upgrade them:
git -C third_party/mem0 fetch origin
git -C third_party/mem0 checkout origin/main
git add third_party/mem0Then rerun install_codex_plugin.py so the generated local marketplace copy and
hook paths are refreshed.
The default --mcp-transport auto chooses stdio when --env-file is present.
Use --mcp-transport http only when you explicitly want direct HTTP MCP config
with bearer_token_env_var; in that mode, Codex must receive the named token
environment variable before a new thread starts.
Multiple instances can use different plugin IDs and token variables:
printf '%s\n' "$MEM0_HOME_MCP_TOKEN" | \
python3 plugins/mem0-oss/scripts/install_codex_plugin.py \
--name mem0-home \
--display-name "Mem0 Home" \
--url https://mem0-home.example.com:18443/mcp \
--token-stdin \
--token-env-var MEM0_HOME_MCP_TOKEN \
--with-hooks \
--installOpenCode plugin
Releases include a compiled OpenCode package with bundled dependencies and native
tools, hooks and skills. The same install.py entry point handles installation,
updates and rollback while preserving external connection configuration and keys.
See release installation and updates.
OpenCode's official Mem0 plugin is a native TypeScript plugin rather than an MCP-only config. To use that full plugin experience with Mem0 OSS, generate a local OpenCode plugin copy from the official Mem0 submodule and overlay the OSS compatibility client:
git submodule update --init --depth 1 third_party/mem0
printf '%s\n' "$MEM0_OSS_MCP_TOKEN" | \
python3 plugins/mem0-oss/scripts/install_opencode_plugin.py \
--url https://<bridge-host>/mcp \
--token-stdin \
--installThe installer writes a generated copy under
~/.mem0-oss-mcp/opencode-plugins/<name>, builds it with Bun, and installs a
small loader in ~/.config/opencode/plugins/<name>.js. OpenCode loads local
plugins from that directory at startup. The generated plugin keeps the upstream
OpenCode hooks, native tools, and skills, while its memory client forwards
operations to mem0-oss-mcp through JSON-RPC tools/call.
Generated plugins default to explicit capture: periodic message capture is off, and memories are written through the plugin's explicit memory tools. The upstream client-side Dream workflow is also off because durable consolidation belongs on the server. These defaults can be changed deliberately:
--auto-capture-mode boundedsamples every tenth message, with normalized duplicate suppression and per-session/per-day limits.--auto-capture-mode legacyrestores the upstream every-third-message behavior.--client-dreamre-enables the upstream client-side Dream workflow.
Existing MEM0_OSS_AUTO_CAPTURE_MODE and MEM0_DREAM environment values take
precedence over installer defaults.
Token values passed with --token-stdin or --token are written to a local
private dotenv file under ~/.mem0-oss-mcp/opencode-plugins/env/; they are not
written to generated TypeScript source. Pass --env-file when you want to
choose the dotenv path yourself.
To update the upstream OpenCode plugin files, update the third_party/mem0
submodule and rerun install_opencode_plugin.py.
Pi plugin
Releases include prebuilt mem0-oss-pi-<version>.zip, .tar.gz and .sha256
attachments. Extract a package into a persistent directory and run
pi install /path/to/mem0-oss. Set your MCP endpoint and credentials at runtime;
no local build or upstream checkout is required. See
archive installation for Pi and Senpi configuration.
The Publish MCP release workflow builds, installs and tests the extracted
Pi, Codex and OpenCode packages before uploading them. Manual dispatch can add attachments to an existing
release; provenance.json records the generator commit used for that build.
Generate the official Pi 0.3.2+ extension with a self-hosted MCP adapter. Pi uses
Bearer authentication against the bridge's /mcp endpoint. Use a bridge configured
with Sidecar 0.3.13 or later and bridge 0.1.6 or later for cursor traversal,
write idempotency and caller-bound event receipts. Upgrade Sidecar first, then
the bridge, then regenerate the Pi extension and restart or reload the session.
The Core REST adapter shipped in the initial Pi implementation is replaced;
regenerate it using your MCP URL and MCP token.
printf '%s\n' "$MEM0_OSS_MCP_TOKEN" | \
python3 plugins/mem0-oss/scripts/install_pi_plugin.py \
--url http://<mcp-host>:<port>/mcp \
--upstream-plugin-dir /path/to/mem0-checkout \
--token-stdin \
--installThe default upstream source is third_party/mem0; it must include the sibling
integrations/agent-plugin-core directory. Missing patch anchors stop generation
before replacing the active installation. Node.js validates URLs with the same
WHATWG semantics as the runtime. Building requires pnpm; --no-build uses the
host's TypeScript loader and requires the generated package's zod dependency
(pnpm install --prod inside the package).
Packages live in ~/.mem0-oss/pi-plugins/<name>. Tokens stay outside the package
in an owner-only env file. Use --env-file and --token-env-var for an existing
private file. The older --api-key* flag spellings remain aliases for MCP tokens.
Runtime overrides are MEM0_OSS_MCP_URL, MEM0_OSS_MCP_TOKEN and
MEM0_OSS_MCP_TOKEN_ENV_VAR. An endpoint override to a different origin requires
an explicit runtime token; it cannot reuse the installed file token. Credentials,
query strings, fragments and redirects are rejected. Writes are never retried.
--install adds the generated package to settings.json once and preserves
other settings. Existing packages remain as timestamped backups; build or settings
commit failures retain/restore the active package and credentials. --pi-dir
selects both settings and the generated default config directory. Runtime
OMO_CODING_AGENT_DIR, SENPI_CODING_AGENT_DIR, then PI_CODING_AGENT_DIR override
it. For omo native, install with --pi-dir ~/.omo/agent, then restart or /reload.
Remove the previous Mem0 extension from settings before enabling this copy to
avoid duplicate tools and commands. Preserve any unrelated model auth extension.
Identity is MEM0_USER_ID / userId for the user, and MEM0_APP_ID, then the Git
origin's owner-repository name, then the repository/directory basename for the app.
This keeps separate checkouts of the same remote in the same project scope.
mem0-config.json retains defaultScope, contextInjection, searchThreshold
and autoCapture. Capture defaults to false. Explicit saves use infer=false;
opt-in automatic capture is redacted and marked metadata.type=auto_capture.
The tool accepts metadata on add/update and string filters on search/list and
ID operations. Extra filters cannot override the selected user/app/run scope;
filtered bulk deletion supports the bridge's entity and type filters only.
Select /mem0-scope global explicitly before requesting global tool scope.
The system prompt contains a fixed conservative memory policy. Variable recall
is a hidden mem0-recall message. Senpi's previewSafe hook returns the same
system prefix during prewarm without fetching memories or consuming recall state;
upstream Pi ignores the optional registration flag. Both runtimes are exercised
in CI. This does not establish a particular provider's cache hit rate or cost.
get_all returns one cursor page (20 rows by default, page_size up to 100),
with nextCursor for explicit continuation. The client iterateAllPages method
supports complete traversal beyond 5000 records without hydrating earlier pages
again. /mem0-status counts the active Sidecar index without fetching Core rows;
this index count may exceed currently readable rows if stale records remain.
/mem0-tour shows a bounded preview in the interactive UI; headless mode returns
only a short summary. Memory bodies from tour do not enter model context.
Older servers fail promptly with upgrade guidance rather than falling back to
numeric pagination. The existing numeric MCP/dashboard listing remains unchanged.
ID mutations first read that ID and verify its scope. Unfiltered bulk deletion uses the bridge's scoped endpoint; narrowed bulk deletion reads and deletes one cursor page at a time. Cancellation or failure reports confirmed deletions and does not retry writes. Add returns the durable event ID. Cloud-only custom categories and rerank/source body options are ignored.
Run source/build/runtime verification with tests/run_pi_qa.sh (Node 24, pnpm,
Bun and the pinned submodule). Production data and daily extension settings are
not touched by these fixtures.
Set PI_QA_SIDECAR_SOURCE to a Sidecar 0.3.13+ checkout to also run the real
authenticated HTTP/SQLite chain against 8568 fixture memories. CI pins that
checkout and runs the chain as well as the SDK fixtures.
Run
PYTHONPATH=src python3 -m mem0_oss_mcp.serverDocker:
docker build -t mem0-oss-mcp .
docker run --rm -p 8080:8080 --env-file .env mem0-oss-mcpTools
add_memorysearch_memoriesget_memoriesget_memoryupdate_memorydelete_memorydelete_all_memoriesdelete_entitieslist_entitieslist_eventsget_event_status
list_events and get_event_status are implemented locally because the OSS
REST server writes synchronously and does not expose the platform event API.
This server cannot be deployed
Maintenance
Related MCP Connectors
- KogniteOAuthdev.kognite
Hosted agent memory: store, search, and recall facts across sessions from any MCP client.
An MCP memory server. One memory your agents share — across models, devices and apps.
Cross-vendor AI memory over MCP. One semantic store, readable and writeable from every MCP client.
Personal knowledge graph as an AI memory layer over MCP - read, save, and link your memories.
Related MCP Servers
- AlicenseAqualityNot gradedmaintenanceProvides long-term memory capabilities for MCP clients by wrapping the Mem0 API, enabling semantic search, storage, retrieval, and management of conversation memories across users and agents.9Apache 2.0
- AlicenseNot gradedqualityNot gradedmaintenanceIntegrates the Mem0 Memory API with MCP-compatible clients to provide AI agents with persistent, long-term memory capabilities. It enables users to add, search, update, and delete memories to maintain context and personalization across different interactions.Apache 2.0
- AlicenseNot gradedqualityDmaintenanceAn MCP server that provides memory management tools for AI agents, enabling adding, searching, updating, and deleting memories via the Mem0 API.11 npmMIT
- FlicenseNot gradedqualityDmaintenanceA thin stdio MCP bridge to a self-hosted mem0 OSS server for agent memory, enabling persistent memory operations like adding, searching, and managing durable facts with automatic reconciliation and question-rewrite capabilities.1-