OpenClaw’s memory is unreliable, and you don’t know when it will break

Original Article ↗ Hacker News Discussion ↗

Reliability and Memory Limits

  • Many comments report OpenClaw as brittle and unstable: configs break between releases, it edits its own config incorrectly, blows context windows, and gets “stuck.”
  • Frequent updates are seen as poorly tested, with weak documentation and noisy changelogs; some users freeze versions and hot‑patch bugs.
  • Core pain point: long‑term memory. Retrieval is viewed as relatively easy; deciding what to store, how to summarize, and when to forget is hard.
  • Users try workarounds: repo maps and recursive summaries, markdown “diaries,” RAG systems, hierarchical markdown/semantic DBs, belief-based memory layers, and distributed memory per agent/automation. None are seen as a complete solution.

Use Cases and Perceived Benefits

  • Reported uses include: sales research, proposal drafting, ops staging, landing pages, internal process documentation, daily briefings, reporting, habit tracking, health logging, to‑dos, project management, scheduling, news digests, marketing asset generation, social posting, bug triage, basic SWE/SRE work, and personal development nudges.
  • Some value the “persistent assistant” aspect: shared memory across Telegram/Discord/Slack channels, cron‑triggered tasks, proactive reminders, and cross‑tool orchestration without writing separate scripts.
  • Advocates frame it as process augmentation rather than full automation, emphasizing convenience over novelty.

Skepticism and Critiques

  • Strong view that almost all use cases can be handled as well or better by:
    • Direct LLM use (ChatGPT/Claude/etc.),
    • Simple scripts plus cron,
    • Existing apps (todo managers, CRMs, email clients).
  • Critics argue OpenClaw adds indirection, cost, and unreliability, amounting to “executive cosplay” rather than real productivity.
  • Security concerns: broad access to files, APIs, and messaging; risk of exposing credentials; OpenClaw described as a “security nightmare” or ideal tool for spam/scams.
  • Some see it as wasteful of subsidized tokens and lacking a unique must‑have capability.

Alternatives and DIY Approaches

  • Multiple users describe rolling their own agents using Claude Code/Codex, shell scripts, Telegram bots, orchestration tools (e.g., n8n, windmill), or custom Go/Node/VS Code setups.
  • Competing frameworks and homegrown systems aim for tighter control, simpler memory strategies, and better reliability.

Deeper AI Memory Discussion

  • Several comments generalize beyond OpenClaw: current LLMs are likened to “geniuses with anterograde amnesia.”
  • There is broad pessimism that ad‑hoc RAG and external stores can ever fully match the capability of in‑model, weight‑based long‑term learning with today’s techniques.
  • Some view improved, brain‑inspired continuous learning as a critical open research area; timing of breakthroughs is considered unclear.