mirror of
https://github.com/tiennm99/goclaw.git
synced 2026-07-12 23:05:39 +00:00
viettranx
2731f99ad5
Auto-inject previously searched episodic memory using only the latest user message. Follow-up questions like "what's my favorite?" returned poor matches because the embedding lost the conversational frame. InjectParams now carries an optional RecentContext field that pgAuto Injector prepends to the search query as "Context: ... \nQuery: ..." before running the FTS+vector hybrid search. The "Context:"/"Query:" framing works with both instruction-tuned embedding models (which respect the labels) and plain models (neutral separators). ContextStage walks the message history backward, collects up to 2 trailing user turns capped at 300 runes total, and threads the snippet through the AutoInject callback to the injector. Empty RecentContext preserves legacy single-message search semantics — zero-risk fallback for callers that haven't adopted the new field. Rune-based truncation (not byte) keeps vi/zh locales safe: a byte-wise tail-clip would slice multi-byte runes and emit invalid UTF-8 to the embedding model, degrading exactly the cases Phase 9 is meant to fix. tailClipRunes helper covers Vietnamese, Chinese, Japanese, emoji. 13 regression tests: recall query builder (unicode-safe clip, whitespace handling, position ordering), buildRecentContext (order preservation, turn cap, truncation, non-user skip), and tailClipRunes (CJK, short input, zero cap). All passing with -race. Refs plans/260410-1009-openclaw-ts-feature-port/phase-09-active- memory-recall.md — minimal-viable delivery; Tier 2 LLM re-ranking and per-session recall cache deferred until operational data shows context-aware search alone is insufficient.