#30 First Dates with OpenClaw Memory

1 messages · Page 1 of 1 (latest)

storm flower
#

Hey all, I’m looking for a sanity check on my OpenClaw memory/context setup.

I’m about four weeks in. Early on I trusted my main agent to guide its own setup, which led to a lot of noisy crons, chaos in my Supabase task system, unexpected costs, and still no reliable continuity. Since then I’ve been through several additional conductor-led “memory fixes,” including transcript exports + SQLite, QMD ideas, and most recently LanceDB + Lossless Claw.

The main issue for my conductor has been poor continuity: stale config references, old files being treated as current due to multiple changes, hallucinated state, and even his impersonating another agent. Exasperated, early this morning I finally did a major workspace cleanup and a partial reset of the conductor agent while preserving tools and my secondary agents.

Current reported state:

  • lossless-claw is set as plugins.slots.contextEngine
  • memory-lancedb is set as plugins.slots.memory
  • memory-core is disabled
  • autoRecall / autoCapture are enabled
  • compaction safeguard + memory flush are enabled
  • lcm.db is about 102 MB
  • memory/lancedb/ is only about 380 KB because it was freshly initialized today
  • bootstrapMaxChars: 40,000 / bootstrapTotalMaxChars: 120,000
  • prior QMD-based memory was archived under memory/_archive/

My ask: does this look like the right general architecture for OpenClaw continuity/context, or are there obvious red flags?

I’m especially trying to understand:

  1. Whether Lossless Claw + LanceDB is the right pairing.
  2. Whether disabling memory-core is expected in this setup.
  3. Whether this memory setup sounds sane or overbuilt.

I’m less than 24 hours into this setup and am afraid to start investing my time (again) in rebuilding context. Appreciate any guidance.

tired tulipBOT
#

Sanity check: the shape is reasonable, but I’d be cautious about treating it as “continuity is solved” yet.

  1. Lossless Claw + LanceDB is a valid pairing in principle. They occupy different slots: plugins.slots.contextEngine controls context assembly/compaction, while plugins.slots.memory controls memory tools/recall. Current docs explicitly show lossless-claw as a context engine and memory-lancedb as an install-on-demand long-term memory plugin.

  2. Using memory-lancedb instead of memory-core is expected if LanceDB owns the memory slot. The slot is exclusive: default is memory-core, but setting plugins.slots.memory = "memory-lancedb" means LanceDB is the active memory plugin. I would not describe this as “memory-core disabled” unless you actually set plugins.entries.memory-core.enabled: false; that distinction matters because some memory/dreaming plumbing historically references memory-core. If you have dreaming config, put/verify it under the active memory plugin config if LanceDB owns the slot.

  3. The setup sounds sane, maybe slightly overbuilt — but not absurd. The big red flags are operational, not architectural:

  • Don’t let archived QMD notes remain in places the agent may casually read as current. memory/_archive/ is fine only if your workspace instructions clearly say it is historical/non-authoritative.
  • memory/lancedb/ being tiny after fresh init is expected. It won’t magically contain your old context unless you intentionally import/store it.
  • Auto-capture is not a substitute for a curated source of truth. LanceDB auto-capture is heuristic/trigger-based and short-message oriented; important project state should still live in clean MEMORY.md, AGENTS.md, or explicit docs.
  • Bootstrap limits of 40k / 120k are high-ish. Not automatically wrong, but if stale files are the problem, “more bootstrap” can make hallucinated continuity worse by injecting too much mixed-era context.

My recommendation: run this setup for a

#

few days, but rebuild context deliberately: one concise current-state doc, one durable MEMORY.md, and archive old experiments with loud “historical only” labels. Then use /context detail, openclaw plugins list --enabled, openclaw memory status / LanceDB stats, and openclaw doctor to confirm what is actually loaded.

Docs worth checking:
https://docs.openclaw.ai/concepts/context-engine
https://docs.openclaw.ai/concepts/memory
https://docs.openclaw.ai/tools/plugin
https://docs.openclaw.ai/reference/token-use

Also: if you want deeper troubleshooting here, please include openclaw --version and the relevant redacted plugins config block. Version mismatches matter a lot with this stuff. 🦞

storm flower
#

Thanks! I did a deeper cleanup to remove confusion points. Running 2026.5.5 and hoping the new stack holds without my conductor getting sketchy again. Bootstrap trimmed for added headroom. Thanks again!