Memory & Context Settings: A Complete Guide to Agent Memory on Bot It Out
By Bot It Out Team
Your agents now have a single place to configure all memory and context management features. The new Memory & Context section lives in your instance's Agents tab and puts five powerful controls at your fingertips.
Here's what each one does and when to use it.
Active Memory (Recommended)
What it does: Before your agent replies to any message, a dedicated sub-agent scans past conversations for relevant context and injects it into the prompt. Your agent remembers what you talked about last week without you reminding it.
When to enable: Always. This is the single most impactful memory feature for most users. It's the difference between an agent with amnesia and one that builds a relationship with each user over time.
How it works: The Active Memory plugin runs a bounded, blocking memory search before each reply. It looks at the incoming message, queries the memory index for related past interactions, and prepends the most relevant context. The agent sees this as "things I already know about this person/topic."
Added in: OpenClaw 2026.4.10
QMD (Hybrid Search)
What it does: QMD (Quick Memory Dump) provides BM25 keyword search plus vector similarity search across your agent's knowledge files. It's the engine that makes memory searchable.
When to enable: If you're using Active Memory, you typically don't need QMD separately — Active Memory handles the search. Enable QMD if you want agents to have access to a large knowledge base of documents (not just conversation history) and you're not using Active Memory.
How it works: QMD runs locally on your instance as a sidecar process. It indexes files in your agent's memory directory, builds both a full-text search index (BM25) and a vector index, then answers search queries by combining both result sets. No external API calls, no extra cost.
Note: If you enable both Active Memory and QMD, you'll see a warning — they serve overlapping purposes. Pick one.
Lossless-Claw (LCM)
What it does: LCM ensures your agents never lose context from long conversations. Standard LLM conversations hit a context window limit and older messages get dropped. LCM uses a DAG (directed acyclic graph) structure to persist the full conversation history and reconstruct relevant context on demand.
When to enable: If your agents have long-running conversations (customer support threads, ongoing projects, multi-day negotiations). Less important for short Q&A interactions.
How it works: LCM is a separate plugin that intercepts the conversation management layer. Instead of a flat message list that gets truncated, it maintains a graph of conversation nodes with relationships. When the context window fills up, it intelligently selects which nodes to keep based on relevance to the current topic, not just recency.
First install is slow: The first time you enable LCM, it downloads and installs the plugin package (~1 minute). Subsequent toggles are fast.
ContextEngine
What it does: Controls how your agents prune and compact their conversation context. Without ContextEngine, the default behavior is basic truncation. With it, you get intelligent context management with configurable strategies.
When to enable: For any agent that has conversations longer than a few messages. The default strategy works well for most cases.
Strategies explained:
| Strategy | Cache TTL | Compaction | Best for |
|---|---|---|---|
| Default | 1 hour | Safeguard | Most users — balanced between memory and performance |
| Aggressive | 15 minutes | Auto | High-volume agents with many short conversations |
| Conservative | 4 hours | Safeguard | Agents handling complex, multi-step tasks |
| Full History | No pruning | Safeguard | Debugging or when you need complete conversation records |
Memory Citations
What it does: When your agent recalls information from memory, citations show the source — which conversation or document the information came from.
Options:
- Auto (recommended): Shows citations when the agent draws from memory, hides them for direct responses
- Always: Every memory-assisted response includes source citations
- Never: Citations are suppressed entirely
When to change from Auto: Switch to "Always" if you need auditability (compliance, legal). Switch to "Never" if citations clutter the user experience in your use case.
The Recommended Setup
For most users, we recommend:
| Setting | Value | Why |
|---|---|---|
| Active Memory | Enabled | Smart, AI-powered recall across conversations |
| QMD | Disabled | Redundant with Active Memory |
| Lossless-Claw | Enabled | Prevents context loss in long conversations |
| ContextEngine | Enabled (Default) | Intelligent context window management |
| Citations | Auto | Shows sources when relevant, clean otherwise |
This gives your agents persistent memory, conversation durability, and smart context management — with minimal configuration.
Important Notes
- Toggling any setting restarts the gateway: Your agent will be briefly unavailable (5-10 seconds) while the gateway restarts with the new configuration.
- Active Memory requires OpenClaw 2026.4.10+: If your instance is on an older version, the toggle will be disabled with an upgrade prompt.
- LCM first install takes ~1 minute: It downloads the plugin package on first enable. Subsequent toggles are instant.
- Settings are per-instance: Each instance has its own memory configuration. Different instances can use different setups.
Configure once, and your agents get smarter with every conversation.