Coexist with lossless-claw, or move to Remnic LCM.

Different slots, no conflict

OpenClaw plugins register against slots. lossless-claw occupies the contextEngine slot — it replaces OpenClaw's built-in conversation compaction. Remnic occupies the memory slot — it intercepts before_agent_start and agent_end hooks to inject recalled facts and buffer turns for extraction. The two surfaces are orthogonal.

Run both side-by-side

Useful when you want lossless-claw to keep handling compaction while Remnic builds up extracted-fact memory. A single OpenClaw config sets both slots:

{
  "plugins": {
    "slots": {
      "memory": "openclaw-remnic",
      "contextEngine": "lossless-claw"
    }
  }
}

Each subsystem reads and writes its own SQLite store. They never share storage and never race for the same hook surface.

Migrate to Remnic LCM mode

Remnic ships its own LCM mode: a SQLite archive of every turn plus a hierarchical summary DAG, exposed through Remnic's recall pipeline. The schemas are near-isomorphic, so migration is a direct SQLite→SQLite transform — not a lossy distillation.

1. Enable Remnic LCM mode

{
  "plugins": {
    "entries": {
      "openclaw-remnic": {
        "config": {
          "lcmEnabled": true
        }
      }
    }
  }
}

2. Install the importer

npm install -g @remnic/import-lossless-claw
# or
pnpm add @remnic/import-lossless-claw

3. Run the import

# Preview without writing
remnic import-lossless-claw --src ~/.openclaw/lcm.db --dry-run

# Real run
remnic import-lossless-claw --src ~/.openclaw/lcm.db

# Restrict to specific resolved sessions
remnic import-lossless-claw --src ~/.openclaw/lcm.db \
  --session-filter sess-A --session-filter sess-B

The importer is idempotent: re-running it inserts zero new rows for already-imported turns. The source database is opened read-only with fileMustExist: true, so a half-baked source can never be mutated mid-import.

4. Disable lossless-claw (optional)

Once the import succeeds, drop contextEngine: "lossless-claw" from plugins.slots. OpenClaw falls back to its built-in compaction; Remnic LCM is designed to complement, not replace, that.

What carries over

What is lossy

• Multi-parent summary DAG. lossless-claw can attach one summary node to multiple parents. Remnic LCM is single-parent. The importer keeps the lowest-ordinal parent and reports the collapse count. With the default summaryRollupFanIn = 4 this is rare in practice.
• Message parts. lossless-claw's message_parts table — fine-grained tool I/O, patches, file references, step-start/finish markers — has no Remnic LCM analog. Only the rendered messages.content survives.
• Large files + exploration summaries. Run those through Remnic's normal extraction pipeline first if you want them in facts/; they aren't part of the LCM transform.
• Compaction telemetry. Prompt-cache state and activity bands are owned by lossless-claw's compaction loop, not portable to Remnic.

Import-boundary marker

For each session that gains data, the importer writes one row to lcm_compaction_events with tokens_before == tokens_after. That equality encodes "import boundary, not a real compaction event" — Remnic's own compaction telemetry will start from this anchor on the next run. Token totals come from SUM(token_count) on the destination at boundary-write time, so a partial-retry where only summaries land new this run still records the correct anchor.

Idempotency

Dedup is keyed on source identity (metadata.conversation_id + metadata.source_seq) via SQLite json_extract, with an in-memory pre-fetch per session for O(1) lookup. Two source conversations sharing a single session_id both contribute messages without colliding — turn_index is assigned as a session-global running counter in chronological (not UUID) order.