AI Agents Have a Memory Problem

You've probably experienced this: you tell an AI assistant something, and two conversations later it's completely forgotten. Your preferences, your project context, that thing you spent ten minutes explaining — gone. It's like talking to someone with amnesia every morning.

Most AI memory systems sit there passively, waiting for you to remember to use them. You have to say "remember this" or "search my notes." By then, the moment where memory would have made the reply feel natural has already passed.

Active-Memory solves that.

What Is Active-Memory?

Active-Memory is a plugin that gives every SuperClaws agent semantic memory — the ability to search and recall information from past conversations using meaning, not just keywords.

Unlike basic keyword search, semantic memory understands intent. If you previously talked about your "evening coding routine," a search for "when do I write code?" will find it. The system uses embedding-based vector search combined with traditional keyword matching to surface the right context at the right time.

And here's the key part: it's enabled on every SuperClaws agent by default. You don't need to configure anything, install anything, or opt in. It just works.

How It Works

Active-Memory runs a small, dedicated sub-agent before your agent generates each reply. Here's the flow:

  1. You send a message to your agent.
  2. Active-Memory silently builds a search query from your message and recent conversation.
  3. It searches your stored memories — things like preferences, project details, personal facts, and past decisions.
  4. If it finds something relevant, it injects that context into the agent's prompt, hidden from view.
  5. Your agent replies — now with the benefit of your history.

This all happens in the background. No extra commands, no waiting. Most users never see it running. They just notice that their agent remembers.

What Gets Stored

Memories are organized in Markdown files in your workspace:

  • MEMORY.md — the main, evergreen memory file
  • memory/*.md — additional memory files for different topics
  • Session transcripts — conversations can optionally be indexed for recall

These files are yours. They live in your workspace, not on some external server. You can read them, edit them, or organize them however you want.

Searching Memories

Agents query memories through two built-in tools:

  • memory_search — finds relevant notes using a mix of semantic (meaning-based) and keyword search, then merges and ranks the results
  • memory_get — retrieves specific lines from a memory file when the agent knows exactly what it needs

The search uses a hybrid approach: vector embeddings catch conceptual matches, while BM25 keyword search catches exact terms like IDs, error strings, or config keys. Both run in parallel, and the results are intelligently merged.

Why It Matters

Continuity

Your agent picks up where it left off. If you discussed your tech stack last week, it's still there. If you mentioned your timezone once, your agent uses it forever. This is the difference between an agent that feels like a colleague and one that feels like a search engine.

Personalization

Over time, your agent builds a picture of how you work, what you care about, and what you prefer. It learns your writing style, your project priorities, even your quirks. All without you having to manually configure a "profile."

Reduced Repetition

No more re-explaining the same context. No more typing "like I mentioned before" and hoping the agent finds it. Active-Memory handles the recall so you can focus on the conversation.

Under the Hood

For the technically curious, here's how the pieces fit together:

  • Plugin architecture: Active-Memory is an OpenClaw plugin that registers a blocking sub-agent in the reply pipeline. It runs before the main model generates a response.
  • Embedding providers: Memory search supports OpenAI, Gemini, Voyage, Mistral, Bedrock, Ollama, and local GGUF models. No API key is needed for local or Ollama setups.
  • Hybrid retrieval: Every search runs vector search (semantic similarity) and BM25 (keyword matching) in parallel, then merges results with configurable weighting.
  • Temporal decay: Older memories gradually lose ranking weight (30-day half-life by default) so recent information surfaces first. Evergreen files like MEMORY.md are never decayed.
  • Configurable behavior: You can tune query modes (message-only, recent context, or full conversation), prompt styles (strict to recall-heavy), and timeouts to balance speed vs. recall depth.
  • Session scoping: By default, Active-Memory runs in direct-message sessions. Group and channel sessions can be opted in.

Try It

If you're using SuperClaws, Active-Memory is already running. You can see it in action by typing /verbose on in any conversation — you'll see debug lines showing when memory was found and injected.

Want to check or toggle it? Use these commands:

  • /active-memory status — see if it's on for this session
  • /active-memory off — pause it for this conversation
  • /active-memory on — turn it back on

And if you're not using SuperClaws yet? This is the kind of thing that changes how you interact with AI. Once your agent actually remembers you, going back to stateless chatbots feels like a downgrade.

Try it with your SuperClaws agent today.