Slowave

Slowave is an open-source, locally-running memory layer for AI coding agents, released under the AGPL-3.0-or-later license. It gives tools like Claude Code, Cursor, Cline, Codex, Windsurf, OpenCode, and Claude Desktop a single shared memory that persists across sessions and clients, without sending data to a hosted service or requiring a separate LLM API key. The project is currently in public beta (as noted in the README), with the latest release being v0.20.3 published in September 2026.

What It Is

Slowave is a brain-inspired adaptive memory system that sits between your AI coding agents and their context. Rather than acting as a static RAG store or a Markdown-file manager, it implements a continuous feedback loop — remember → recall → use → feedback → reinforce/weaken → decay — so that useful memories strengthen over time and stale ones lose influence. The core runs entirely on local embeddings and deterministic mechanics, meaning no extra model calls and no LLM API key are needed for memory maintenance.

How the Memory Loop Works

Slowave exposes five MCP tools that agents call during a task:

• Activate — starts a session and loads scoped, relevant memory for the current task and goal.
• Remember — saves a durable fact, decision, preference, or constraint.
• Recall — performs a mid-task memory lookup when the agent needs something specific.
• Feedback — marks retrieved memory as useful, irrelevant, or stale, which adjusts future salience.
• Commit — saves the task outcome and any reusable multi-step procedure.

A background consolidation worker processes raw events into episodes, prototypes, schemas, and relations stored in a local SQLite database. The agent's feedback — not a separate LLM judge — shapes what Slowave returns over time.

Local-First Architecture and Inspectability

All memory is stored in plaintext in the current OS user's application-data directory in an open SQLite database. Slowave does not send memory to a hosted service. A local dashboard (launched with slowave dashboard) lets users browse saved memories, procedures, retrievals, activity logs, a memory graph, and system health. The default embedding model (~45 MB) downloads from Hugging Face on first use and is cached locally for offline operation afterward.

Setup Path

Installation is a two-command process via pipx:

pipx install slowave
slowave setup

The setup command auto-detects supported AI clients on the machine, configures MCP connections and lifecycle instructions, and starts the local daemon and consolidation worker. Per-client setup is also available (e.g., slowave setup --client cursor). Claude Desktop and Cursor require one manual paste step after setup. The README notes that client coverage is actively expanding.

Update: v0.20.3

The latest release, slowave-v0.20.3, was published on September 14, 2026. The GitHub repository was created in June 2026 and has seen active development since, with the last push matching the latest release date. The README explicitly flags the project as public beta software, noting that APIs, configuration, and storage schema may change and that migrations are not guaranteed before stable release. Evaluation notes in the repository cover preliminary retrieval-evidence results with stated methodology and limitations, and the project discourages treating any benchmark result as a production-quality claim.