CausaMem

What CausaMem does

CausaMem, published as MHH-Causality-Memory, is a long-term memory system for AI agents organized as eight layers plus an intuition step: reality evidence, atomic facts, refined summaries, profile scenes, wiki knowledge, dream-style consolidation, causal chains and intuition injection. The aim stated in the README is to let an agent keep roughly two million characters of history and reason over why events happened, not just retrieve similar text.

Version 0.18 narrows the design to a time, person, event, cause and result record for each memory, with reasoning operations such as tracing causes, predicting consequences, separating correlation from causation and checking whether a causal link is still valid. In the OpenClaw integration it injects a cognitive anchor in before_prompt_build and captures conversations in agent_end, with a deterministic gate committing only approved memories. The README is in Chinese with translations linked, and the author reports a 96.7% pass rate on a 30-question pilot.

Key features

• Eight layers from raw evidence to causal chains and intuition injection
• Time, person, event, cause and result structure for each memory
• Thirteen causal reasoning operations including intervention and multi-cause attribution
• OpenClaw hooks for prompt injection and conversation capture
• Deterministic gate that commits only approved memories
• README translations in more than 30 languages

When to use it

• Giving an agent a durable memory that records why decisions were made
• Checking whether an old rule or decision still applies
• Experimenting with causal memory structures for long-lived agents

Who it is for: Developers building long-lived agents who want causal, traceable memory rather than plain similarity search.

How it fits with Hermes Agent

Its repository description and topics label it an OpenClaw and Hermes memory project, but the integration the README details is for OpenClaw.

FAQ