OpenClaw Releases Markdown-Based Agent Memory

OpenClaw has released an agent memory system composed of Markdown files and a lightweight index (MEMORY.md), without complex databases or heavy frameworks. This system has been validated in a production environment for 7 months, based on the insight that reliability can suffer if write/read paths do not reflect temporal context, even if the database schema accounts for it.

Architecture and Core Principles

Memory is managed via Markdown files within a Git repository, with SQLite indexes and semantic search used only as auxiliary tools to improve lookup speed. The files themselves serve as the source of truth, taking precedence over the index in case of discrepancies. At boot, only the Identity layer and MEMORY.md (routing index) are loaded, while the rest are retrieved from the narrowest scope of files during queries.

Write and Read Path Rules

To ensure reliability, the following strict editing rules are applied.

• Provenance Tag: Sources are explicitly marked with tags such as [stated], [observed], and [inferred] to prevent agent suggestions from being mistaken for user decisions.
• Recurrence Gate: Inferred lessons are stored only after 3 or more independent signals are confirmed across 2 or more sessions, and signals older than 30 days have their weight halved.
• Supersession: When new information arrives, existing lines are not deleted; instead, strikethroughs and dates are added to preserve history. Validity periods and sources can be tracked via Git log/blame.
• Read Path: Memory retrieval is mandatory before answering, returning a compact bundle limited to 5 sources by default. Each fact includes the file path, line number, source, and freshness.

Validation Results and Limitations

An existing memory runtime with 382 dependencies returned past deprecated information (GraphQL) as current, whereas the Markdown system accurately returned current facts (tRPC) and their sources. This system works with 3 frontier models from 2 vendors, maintaining the agent's identity, preferences, and decisions even when models are swapped. However, LLM authors must consistently apply editing discipline, and actual locking and merge discipline are required in large team environments.