Memory Layer

Memory Layer is a local-first memory system for coding agents and developers. It turns project work into durable, searchable knowledge, so the next Codex, Claude, or human session can start with evidence instead of guesswork.

It captures what happened, curates what matters, stores it in PostgreSQL with pgvector, and exposes it through a TUI, browser UI, and agent-friendly CLI.

Website · Documentation

Start here

Try it with Docker

The reproducible demo needs only Docker; it starts PostgreSQL with pgvector, the service, and the web UI. Clone the repository and start the stack:

git clone https://github.com/3vilM33pl3/memory
cd memory
docker compose up

In a second terminal, load the demo project and ask one question:

docker compose exec memory memory demo
docker compose exec memory memory query --project demo --question "How does reinforcement work?"

Open http://localhost:4040 for the browser UI, or run docker compose exec memory memory tui for the terminal UI. Follow the Quickstart for the complete demo path.

Use it in a project

For a native installation, prerequisites, and verification, start with the Install guide or download a package from GitHub Releases. The setup wizard configures the machine once and the project you are working in:

memory wizard --global
cd /path/to/your-project
memory wizard --dry-run
memory wizard
memory doctor

Native installations need a PostgreSQL database with pgvector; the install guide covers local, hosted, Windows, and package-specific paths. The Windows x86_64 MSI is per-user, installs under %LOCALAPPDATA%\Programs\Memory Layer, and adds its bin directory to the user PATH.

Work with it every day

Capture finished work, query before making a change, resume after an interruption, and inspect or recover when something looks wrong. The Daily workflow gives the commands and the TUI and Web UI make the stored evidence easy to inspect.

Connect agents

• Codex Desktop plugin — the supported MCP connection and desktop workflow skill.
• Agents — project setup and agent-facing workflows.
• MCP — read-only project-memory tools over stdio or local Streamable HTTP.

Use one Memory Layer MCP connection per client. The Codex plugin guide explains how to avoid duplicate tool registration and how to verify the active project.

What it provides

• Cited answers from lexical, semantic, relation, and code-graph retrieval.
• Project memories with provenance, curation, review proposals, and durable re-entry briefings.
• Local TUI and browser UI for memories, activity, review, watchers, and runtime health.
• Coding-agent integration through the CLI, repo-local skills, and MCP.
• Repeatable evaluation with paired ablations, immutable artifacts, gates, and cost/latency reporting.

Evidence and evaluation

The newest checked-in local reference is the 2026-07-06 memory-quality-v1 re-baseline: 0.692 (18/26), with retrieval at 10/10 and grounded answers at 8/9. Its gate is still red because the adversarial-stale floor is intentionally unmet; treat it as a precise engineering reference, not a general release claim.

The separately recorded 2026-05-03 Docker memory-improvement-v1 benchmark ran five paired repeats against a different suite. It reported full-memory aggregate success from 0.0% to 18.1%, perfect retrieval ranking metrics, and a 41.2% reduction in total tokens. Read it as historical Docker-harness evidence rather than a directly comparable successor to the July local run.

For methodology and commands, see the evaluation guide and CLI reference.

Documentation

The public site is organised with an essentials-first path and optional deep dives:

• Quickstart
• Install
• Daily workflow
• Help
• Operations
• How it works
• CLI reference

The docs-site/ directory contains the public site; the docs/ tree contains the detailed in-repository manual and developer reference.

Development and contributing

Start with Contributing, the developer documentation, and the dev-stack guide. The development stack is isolated from packaged installations; its setup and verification steps belong in that guide rather than this README.

License

Memory Layer is dual-licensed:

• Open source: GNU Affero General Public License v3.0 or later; see LICENSE.
• Commercial: available under a separate commercial license; see LICENSE-COMMERCIAL.md.

Contributions are accepted under the repository's open-source license unless explicitly agreed otherwise in writing. See Contributing.