membrane— agent skill

<p align="center"> <img src="static/img/logo-mark.svg" width="72" height="72" alt="Membrane logo"> </p>

<h1 align="center">Membrane</h1>

<p align="center"> <strong>Memory that agents can connect, revise, and learn from.</strong> </p>

<p align="center"> <a href="https://membrane.gustycube.com">Documentation</a> · <a href="https://membrane.gustycube.com/quickstart">Quickstart</a> · <a href="https://membrane.gustycube.com/architecture">Architecture</a> · <a href="https://membrane.gustycube.com/api/overview">API reference</a> </p>

Membrane gives LLM agents persistent, structured memory. An agent can remember what happened, connect it to existing knowledge, retrieve the context a task needs, and correct that knowledge when circumstances change.

Use it for agents that carry work across sessions, maintain changing facts, reuse successful procedures, or need to explain which memories informed an answer.

What changes when an agent has memory?

• Context stays connected. Facts, episodes, procedures, and plans link through

shared entities. Retrieval returns a bounded graph of relevant records.

• Knowledge can be corrected. Supersede stale facts, contest uncertain claims,

fork alternatives, or retract a record while retaining its audit history.

• Experience influences future work. Reinforcement, decay, and consolidation

change what gets remembered and retrieved over time.

• Access follows policy. Scope and sensitivity rules govern reads and writes;

restricted responses omit or redact information the caller cannot access.

How it fits together

flowchart LR
    Experience["Events, observations<br/>and task state"] --> Memory["Typed memory<br/>+ entity links"]
    Memory --> Context["Relevant context<br/>within access limits"]
    Context --> Agent["Agent acts"]
    Agent -->|"New evidence and outcomes"| Memory

Five memory layers, connected by entities:

Explore the memory model →

Choose your integration

For a complete agent loop, see the agent harness. It exercises capture, graph retrieval, revisions, and access controls against a running daemon, with both deterministic and optional live-model scenarios.

Run it locally

[!IMPORTANT] PostgreSQL with pgvector is required, whether you embed the Go library or run the daemon. Embedding and LLM providers are optional: structured capture, graph retrieval, revisions, and access controls work without a model API key.

<details> <summary><strong>Build and start the daemon</strong> — Go, Docker Compose, and Make</summary>

Use the Go toolchain declared in go.mod. From a fresh checkout:

git clone https://github.com/BennettSchwartz/membrane.git
cd membrane
make build

export MEMBRANE_POSTGRES_PASSWORD="$(openssl rand -hex 24)"
export MEMBRANE_POSTGRES_DSN="postgres://membrane:${MEMBRANE_POSTGRES_PASSWORD}@127.0.0.1:5432/membrane_test?sslmode=disable"
docker compose up -d
until docker compose exec -T postgres pg_isready -U membrane -d membrane_test; do sleep 1; done
./bin/membraned --postgres-dsn "$MEMBRANE_POSTGRES_DSN"

The daemon applies its database schema on startup and listens on 127.0.0.1:9090. Its default policy permits the default scope at LOW sensitivity. See configuration for scopes, provider settings, TLS, and authentication.

The TypeScript and Python quickstarts connect to this daemon.

</details>

Add capabilities as you need them

Full quickstart · Deployment guide · Security guide

Explore the project

<details> <summary><strong>Development and verification</strong></summary>

Go uses the toolchain in go.mod. The TypeScript SDK and examples need Node.js 20.19+, the docs toolchain needs Node.js 22+, and the Python SDK needs Python 3.10+.

Set disposable PostgreSQL DSNs to include database integration tests. Provider credentials enable live embedding and model evaluations. See CONTRIBUTING.md and the harness guide.

</details>

Contributing

Bug reports, integration feedback, and contributions are welcome. Open an issue or read the contribution guide.

MIT licensed. See LICENSE.