AI Agent Memory Systems:
A Practical Comparison

Section 1

The Baseline — What OpenClaw Gives You

Out of the box, OpenClaw ships with a solid memory foundation. Most agents can go a long way with just this.

OpenClaw's built-in memory system includes:

• File-based memory — MEMORY.md, daily notes, and a working buffer for in-session state
• SQLite + Vector embeddings — persistent storage with semantic retrieval
• Hybrid search — BM25 (keyword) + vector (semantic) combined for recall
• Optional: MMR re-ranking — Maximal Marginal Relevance to reduce redundancy in results
• Optional: Temporal decay — recency boost that makes recent memories more likely to surface
• Pre-compaction memory flush — structured export of key context before the conversation window is compressed

Built-in Alternatives

OpenClaw uses a plugin slot (plugins.slots.memory) for its memory backend — meaning the whole memory system can be swapped out. Three options exist within the OpenClaw ecosystem:

• memory-core (default) — file-backed, SQLite, manual memory management. The baseline described above.
• memory-lancedb (Community Plugin) — conversation-based memory using LanceDB with automatic capture and recall:
  • Auto-Capture: after every turn, preferences, facts, and decisions are automatically extracted and stored
  • Auto-Recall: before each response, relevant memories are automatically injected into context — no tool call required
  • Install: openclaw plugins install @noncelogic/memory-lancedb
  • Note: autoRecall defaults to false — must be explicitly enabled
• QMD — also lives in the memory slot, but focuses on retrieval improvement (reranking, query expansion) rather than replacing the memory model entirely

Important: memory-core, memory-lancedb, and QMD all occupy the same plugin slot — only one can be active at a time. Lossless Claw, by contrast, sits in the contextEngine slot and is fully combinable with any of the above.

These gaps are the reason third-party memory systems exist. But they solve different problems — which is the key thing to understand.

Section 2

Two Layers of Memory

Before comparing tools, there's a fundamental distinction worth naming clearly.

Memory systems for AI agents operate on two separate layers. Mixing them up leads to picking the wrong tool:

These are complementary, not competing. Infrastructure makes sure all the raw material exists. The meaning layer makes sense of it. You can use one without the other — but both together is more powerful than either alone.

Section 3

The Four Candidates

Four systems are worth knowing about. They sit at different positions on the infrastructure–meaning spectrum and vary widely in integration complexity.

Four systems mapped by memory layer and integration effort. Bottom = infrastructure. Top = meaning & personality.

Section 4

What Works Together?

Not all combinations make sense. Here's what we know about compatibility.

Lossless Claw and Honcho operate on different layers — no conflict, maximum coverage.

Section 5

Decision Guide

Start with a concrete problem, not a technology. Here's how to map your situation to the right tool.

• →
  "Context is getting lost after compaction. I need to fix this."
  Lossless Claw — install it now, zero config required.
• →
  "My agent isn't finding the right memories when it searches."
  QMD — better retrieval, same interface. Use if you're not running Lossless Claw.
• →
  "I want my agent to genuinely understand the user over time."
  Honcho — reasoning over memory, not just retrieval. Needs setup, worth it.
• →
  "I want the agent to develop a real personality, with beliefs and goals."
  cortex-engine — long-term investment. No OpenClaw adapter yet, but the most sophisticated option.
• →
  "I want nothing to be written manually — memory should just happen."
  memory-lancedb — auto-capture and auto-recall out of the box. Enable autoRecall and it runs without any manual memory calls.
• →
  "I'm just getting started and don't have a specific problem yet."
  OpenClaw default is good enough — don't add complexity until you hit a concrete wall.

Section 6

Current Status & What We Haven't Tested

This is a research summary, not a production review. Here's where things actually stand.