polymnemo

polymnemo

[CI] [codecov] [License: MIT]

[Python] [MCP] [Postgres] [Neon] [SQLAlchemy] [Pydantic] [Cloudflare R2] [Azure] [Cloud Run] [Terraform] [Docker] [pytest] [Ruff]

A shared long-term memory across any LLM, over MCP.

Point Claude Desktop, an MCP-capable IDE, or any MCP client at one polymnemo endpoint and they share the same memories — stored in your own Postgres. Store a fact with one assistant, recall it from another; save whole sessions and reload them; even attach files, images, or video. Embeddings run locally (no embedding API key), and the server makes no generative-LLM calls.

Status: active development. Semantic memory + pgvector store, session save/reload, and multimedia memories all work; deployable to Azure Container Apps (or Cloud Run) via Terraform. Wiki · Issues

Features

• 🔗 Cross-LLM shared — point any MCP client at one endpoint; they share the same memory.
• 🧠 Semantic recall — vector search over Postgres + pgvector by default, with exact-text and fused (RRF) modes when you need to find an ID or a name.
• 💬 Sessions — save a full transcript and reload it verbatim, or recall across it.
• 🖼️ Multimedia — attach files, images, or video; bytes go to object storage, only a searchable description is embedded.
• 🔒 Local & private — embeddings run locally (ONNX): no embedding API key, and no generative-LLM calls, ever.
• 👥 Namespaces — "born-shared" collections readable by everyone, vs. private-to-owner; writes are always owner-scoped.
• 🧩 Pluggable layers — store, embedder, auth, retriever, blob store, and rate limiter are all swappable Protocols.
• ☁️ Multi-cloud deploy — one Terraform stack to Azure Container Apps or Cloud Run, scale-to-zero.
• 🚦 Rate limiting — optional global token bucket.

How it works

flowchart LR    Clients["MCP clients<br/>(Claude Desktop, IDEs, …)"] -->|"/mcp · Bearer key"| P["polymnemo<br/>(MCP server)"]    P --> DB[("Postgres + pgvector<br/>text + pointers")]    P -. "large files<br/>(presigned URLs)" .-> OS[("Object storage<br/>S3 / R2")]

A request carries a bearer key (which resolves to a user_id); the tool passes the rate-limit gate, then delegates to a MemoryService that chunks + embeds text and stores the vectors in pgvector — large files go to object storage via presigned URLs, with only a searchable description embedded.

Every layer is a typing.Protocol, wired together by a composition root (context.py), so you can swap an implementation without touching the tools:

The ping tool returns the active layers, so you can see how a running server is wired.

Quickstart

Requires Python 3.11+.

1. Install

python -m venv .venvsource .venv/bin/activate            # Windows: .venv\Scripts\activatepip install -e .                     # add ".[dev]" for the test + lint tooling

2. Provision Postgres (pgvector)

The durable store is Postgres + pgvector; the easiest hosted option is Neon (use the pooled connection string). Apply the schema once:

psql "<your-connection-string>" -f scripts/schema.sql

3. Configure

Copy .env.example to .env and set the database URL and at least one API key:

POLYMNEMO_DATABASE_URL=postgresql://user:pass@host/db?sslmode=requirePOLYMNEMO_API_KEYS=sk-alice-secret:alice,sk-bob-secret:bob   # "key:user_id" pairs

Each key maps a bearer token to a user_id; writes are scoped to that user.

4. Run

polymnemo                            # Streamable HTTP at http://127.0.0.1:8000/mcp

Or with Docker:

docker build -t polymnemo .docker run -e POLYMNEMO_DATABASE_URL="..." -e POLYMNEMO_API_KEYS="sk-alice-secret:alice" \           -e PORT=8080 -p 8080:8080 polymnemo

5. Connect an MCP client

Any client that supports remote (HTTP) MCP servers with custom headers needs two things:

• Endpoint: http://<host>:<port>/mcp
• Header: Authorization: Bearer <your-key>

For clients that read an mcpServers config:

{  "mcpServers": {    "polymnemo": {      "url": "http://127.0.0.1:8000/mcp",      "headers": { "Authorization": "Bearer sk-alice-secret" }    }  }}

Or verify with the inspector:

npx @modelcontextprotocol/inspector# Transport: Streamable HTTP · URL: http://127.0.0.1:8000/mcp# Header:    Authorization: Bearer sk-alice-secret  → call ping / remember / recall

Concepts

• Users & keys — sign in with GitHub, or present a bearer token. Either way the request resolves to a user_id, and writes are owner-scoped (you can only edit or delete your own memories).
• Headless clients — CI, cron jobs and server-side agents can't complete a browser sign-in, so they use an API token instead. Sign in with GitHub at <server>/tokens to create, list and revoke them; a new token is shown once and stored only as a SHA-256 hash. Deliberately not MCP tools — a secret must never travel an LLM channel.
• Namespaces — memories live in namespaces. A shared namespace (default shared) is readable by everyone ("born shared"); everything else is private to its owner. Sessions and media default to private namespaces.
• Chunking — long content is split into chunks on write (one vector each), so remember may return several ids and recall returns the closest chunks.
• Secret scanning — optional, and off unless you turn it on: set POLYMNEMO_RSP_CONFIG to a scanner config (the image ships one at /app/plugins/rsp.toml). Unset, every write goes through a no-op guard and nothing is scanned. When on, the text of a memory is scanned before it is embedded or stored, so a refused write was never written rather than deleted afterwards. It covers remember, save_session, update and a media memory's description. The contents of uploaded files are not scanned and are stored as you sent them: the bytes go straight to object storage under a presigned URL and never pass through this server. A file's description is scanned; the file is not.

Tools

polymnemo exposes MCP tools for storing, searching, and managing memories:

• Memory — remember, recall, list_memories, get_memory, update, forget

recall takes a mode:

keyword is an exact substring match. rrf fuses the top 50 candidates from each leg, so paging through it reaches at most about 100 results. score is only comparable within one mode, and total is the size of the namespace searched, not the number of matches.

• Sessions — save_session, load_session
• Media — create_upload, confirm_upload, get_download_url

Plus a ping health check and a memory://{namespace} resource for auto-injecting a collection. Media bytes go to object storage via presigned URLs — never through the MCP channel — with only a searchable description embedded (needs the blob extra).

Full arguments and return shapes live in the dedicated MCP tools reference (coming soon). See Concepts for how keys, namespaces, and chunking work.

Configuration

POLYMNEMO_* environment variables (or .env) — full list in .env.example. The essentials:

Development

pip install -e ".[dev]"pytest                               # fast, offline (stub embedder, in-memory store)ruff check . && ruff format --check .   # lint + format (enforced in CI)

Postgres tests run when TEST_DATABASE_URL points at a pgvector Postgres. CI (.github/workflows/ci.yml) runs lint + the suite with coverage and posts a pass/fail/coverage table to each run's summary.

Deploy

polymnemo is stateless (all state in Neon + object storage), so it runs on Azure Container Apps (primary) or Google Cloud Run with scale-to-zero. Everything is Terraform in deploy/terraform/: shared neon/ (Postgres) and r2/ (media bucket) roots own the durable state, and a compute root deploys a service that reads both — so memories and media are shared across clouds.

The Azure path deploys from CI in one click: set the GitHub secrets (scripts/setup-github-secrets.sh), bootstrap the state backend (scripts/bootstrap-tfstate-azure.sh), then run the deploy (azure) workflow (neon → schema → r2 → build → app). GCP is a manual failover. Full walkthrough in docs/deploy.md.

License

MIT — see LICENSE.