Total Recall

Yesterday's chat got compacted away? Per-session event store with FTS + vector recall — pulled back on demand.

Context compaction silently drops the messages that matter most by the time you wish they were still there. ENGRAM catches them on the way out.

Every assistant response is ingested into a per-session event store with full-text search and Ollama-powered vector embeddings (mxbai-embed-large by default). Before each prompt, the most relevant past events are retrieved and injected as a ## Retrieved Memory Context block within a tunable token budget (default 2000). Messages about to be lost to compaction are persisted first via before_compaction, so nothing important disappears. The agent gets a recall tool for explicit memory queries. Pointer compaction keeps storage efficient at scale; null retention means the engram is yours forever.

Install

openclaw plugins install @globalcaos/openclaw-total-recall

Pull the embedding model: ollama pull mxbai-embed-large. Then enable in openclaw.json:

"plugins": {
  "allow": ["tinkerclaw-total-recall"],
  "entries": { "tinkerclaw-total-recall": { "enabled": true } }
}

Pairs Well With

• @globalcaos/openclaw-memory-enhancements — the hippocampus index makes ENGRAM lookups O(1) on known concepts. Big speedup once the corpus grows past a few thousand events.
• @globalcaos/openclaw-identity-persistence — your SOUL.md shapes what's worth remembering. ENGRAM stores the events; CORTEX makes sure recall stays in character.
• @globalcaos/openclaw-round-table — debate traces flow into the same engram. The agent stops rehashing questions the panel already answered last week.

👉 https://github.com/globalcaos/tinkerclaw 👉 https://thetinkerzone.com

Clone it. Fork it. Break it. Make it yours.