Skip to content

Verbatim semantic lane

openclaw-mem now ships a bounded verbatim semantic lane for episodic recall.

The point is narrow on purpose: - improve retrieval over raw episodic evidence - keep the lane read-only - preserve scope / redaction / audit posture - avoid collapsing durable / episodic / working / docs into one blob

This is a retrieval tactic, not a new memory type.

What it is

The lane works over the existing episodic_events.search_text substrate.

Shipped operator surfaces:

openclaw-mem episodes embed --scope <scope> --json
openclaw-mem episodes search "<query>" --scope <scope> --mode hybrid --trace --json
  • episodes embed builds embeddings for the redacted/bounded episodic search text
  • episodes search --mode lexical keeps the existing FTS path
  • episodes search --mode hybrid fuses FTS + vector results with RRF
  • episodes search --mode vector uses vector-only ranking when you want a pure semantic probe

The existing episodes replay and episodes query flows remain unchanged.

What it is not

This slice does not: - create a new durable memory type - auto-promote verbatim hits into L1 durable truth - make Working Set a source corpus - rewrite docs cold-lane retrieval - merge verbatim scores with synthesis/pack scores as if they were directly comparable

If you need durable truth, use the existing L1 store discipline. If you need current-turn activation, use Working Set / pack surfaces. If you need operator-authored docs, use docs cold lane.

Use the verbatim lane when you need questions like: - "which session actually discussed this?" - "what was the raw wording / evidence?" - "find the conversation trail, not just the durable summary"

Typical flow:

# 1) refresh embeddings for the relevant episodic scope
openclaw-mem episodes embed --scope openclaw-mem --limit 500 --json

# 2) semantic-first retrieval over raw episodic evidence
openclaw-mem episodes search "verbatim semantic lane" \
  --scope openclaw-mem \
  --mode hybrid \
  --trace \
  --json

# 3) replay the chosen session when needed
openclaw-mem episodes replay <session_id> --scope openclaw-mem --json

Memory-type × retrieval-lane matrix

memory role lexical / FTS hybrid verbatim semantic dual-language assist graph/topology
durable (L1) yes yes evidence-only yes optional provenance assist
episodic yes yes yes — primary shipped slice later / query-assist only no
working set consume only consume only consume only no direct writeback role optional downstream consumer
docs cold lane yes yes separate docs substrate, not this slice yes optional repo/path routing

Interpretation: - verbatim semantic lane is strongest on episodic evidence recall - durable memory may consume it as supporting evidence, not as automatic truth - working set may later consume the lane, but should not become a source corpus for it - dual-language remains an assistive retrieval booster, not the lane itself

Dual-language relation

The existing dual-language strategy solves query-language mismatch. The verbatim semantic lane solves raw-corpus retrieval over episodic evidence.

Those are related, but not the same thing.

Current shipped posture: - verbatim lane indexes the canonical episodic search_text - --query-en is available as an optional query-side assist for multilingual embedding lookup - this slice does not add a second episodic text_en storage plane

That keeps the first production slice bounded while still leaving room for future cross-language recovery work.

See also: Dual-language memory strategy.

Safety and tradeoffs

Tradeoffs to remember: - stronger semantic retrieval can increase candidate noise - raw episodic evidence can contain transient misunderstandings - vector quality depends on embedding freshness and model consistency

Guardrails in this slice: - scope-aware filtering - redaction-first episodic substrate - read-only retrieval - bounded result counts - optional --trace receipts for lane inspection

  1. Prove lexical episodic search on your scope.
  2. Run episodes embed for that scope.
  3. Compare --mode lexical vs --mode hybrid on a small fixed query set.
  4. Keep route/pack policy unchanged until the semantic lane shows a real hit-quality gain.

If you need the deeper design rationale, see: - Architecture - Mem Engine - Verbatim semantic lane v0 spec (maintainer archive)