Every coding agent session starts from zero. The codebase is there. The documentation is there. But the context behind decisions — why this pattern, why not that one, what failed at 2 AM last Tuesday — is gone. You re-explain it. The agent re-learns it. The cycle repeats.
Most memory systems try to solve this with vector search, RAG, and graphs. They store everything and retrieve by similarity. But after months of 8-hour coding days, the memory grows. Irrelevant entries crowd the context window. A separate LLM layer gets added to summarize and judge relevance — cost, latency, and a split-brain problem where a second model decides what the first model should remember.
Slowave takes a different approach. It makes the agent participate in maintaining its own memory.
The Feedback Loop
Slowave turns each task into a loop: remember → recall → use → feedback → reinforce/weaken → decay.
When the agent retrieves memories, it tells Slowave whether they were useful, irrelevant, or stale. Slowave uses that signal to adapt salience — no separate LLM judge needed. The agent is already evaluating what helped; Slowave handles the mechanical part of adapting the memory substrate.
Retrieval works on this continuous loop of feedback, reinforcement, and decay. Memories that help persist. Memories that don't, fade.
Local-First Architecture
Slowave runs fully locally. A lightweight multilingual embedding model and SQLite — no external memory service, no LLM API required. It exposes 5 MCP endpoints and currently supports Claude Code, Codex, Cursor, Cline, OpenCode, Windsurf, and Claude Desktop.
A local dashboard lets you inspect and delete memories, procedures, retrievals, and see metrics on how valuable the memory layer is for your agents.
Why This Matters for MCP Builders
The MCP integration is the key. Slowave doesn't ask you to adopt a new protocol or rewrite your stack. It plugs into the tools you already use through standard MCP endpoints. The memory layer becomes infrastructure, not another application to manage.
For developers shipping MCP servers and agent workflows, this solves a concrete problem: agent continuity across sessions without the operational overhead of a hosted memory service. The agent remembers what matters because it told the memory layer what mattered.
The project is in public beta. The approach — agent-driven memory maintenance instead of LLM-judged retrieval — is worth watching whether you adopt Slowave specifically or build your own loop.
If this was useful, you can support my open-source work on Ko-fi or check out my services.