Skip to main content

​
Slowave is not a search index with a model on top

Many memory tools work like a library. They store text, fetch the closest matches for a question, and hand them to a language model. This approach, often called RAG, is excellent for searching documents and knowledge bases. Slowave does a different job. It is a living memory that changes as your agent works.
A typical storage and retrieval setup stores text, fetches the closest matches and hands them to a language model, with no learning. Slowave is a loop: the agent writes memories, receives relevant ones, reports what helped, and background upkeep strengthens, fades or replaces memories without an LLM.A typical storage and retrieval setup stores text, fetches the closest matches and hands them to a language model, with no learning. Slowave is a loop: the agent writes memories, receives relevant ones, reports what helped, and background upkeep strengthens, fades or replaces memories without an LLM.

A one-way store-and-fetch pipeline versus a memory that learns from use.

It learns from use

Your agent reports which memories helped. Helpful ones come back more often. Unhelpful ones fade.

It keeps itself current

A newer decision replaces an older one. A plain store would return both and leave your agent to guess.

It captures how, not only what

Besides facts, it saves methods that worked, as procedures.

Its upkeep needs no LLM

Tidying, ranking and recall run locally with no LLM calls. A small local search model handles matching.
Your agent’s own model still does the reasoning. Slowave handles memory. Your agent handles reasoning.
Slowave and document search work well together. Use search for large reference material. Use Slowave for the decisions, rules and lessons that build up as you work.

​
Compared to other ways of giving an agent context

They work well together. Keep stable team rules in a rules file. Let Slowave handle the decisions and lessons that build up as you work.

​
When Slowave is a good fit

  • You use one or more coding agents every day.
  • You are tired of repeating yourself.
  • You want your memory to stay on your own machine.

​
When it is not

  • You need memory for end users of a product you ship. Slowave is for developers using coding agents.
  • You need encrypted storage out of the box.

​
Preliminary results

We test whether Slowave returns the evidence needed to answer questions about long conversation histories. An external LLM judge scores the retrieved text. Scores are lower for time-based questions (about 55% on LoCoMo, 46% on LongMemEval) and for counting across sessions (53%).
These results are preliminary. They measure whether the right evidence was found, not whether a coding agent finished its task better. LongMemEval oracle has no distractor sessions, so it isn’t comparable to LongMemEval-S leaderboards. We do not rank Slowave against other systems.

Read the code

Slowave is open source under AGPL-3.0-or-later. A commercial license is available on request.