#17129: feat(memory): compaction-aware conversation memory with smart-trim

## Summary Adds a **memory-context** plugin that gives agents long-term conversation memory across compaction cycles. When the context window compresses old messages, memory-context archives them into a warm store and recalls relevant context on-demand via hybrid search (BM25 + vector). > **This plugin is disabled by default** and has zero impact on existing functionality. To enable it, add the following to `~/.openclaw/openclaw.json`: > > ```json > { > "plugins": { > "entries": { "memory-context": { "config": {} } }, > "slots": { "memory": "memory-context" } > } > } > ``` > > Both `entries` and `slots` must be present — if either is missing the plugin stays inactive. ## Why This Feature Without long-term memory, agents lose all conversational context once the context window compresses old messages. Users have to repeatedly re-explain background, preferences, and prior decisions. This plugin solves that by: - **Preserving context across compaction** — archived segments survive context-window truncation and are searchable indefinitely - **Automatic recall** — relevant history is injected before each response, so agents "remember" prior conversations without user effort - **Noise-free storage** — heartbeat pings, system prefixes, and junk messages are filtered out, keeping recall clean and relevant - **Zero-config local mode** — works out of the box with a local embedding model; optionally upgrades to Gemini for higher quality semantic search - **Non-blocking** — store loads asynchronously and never delays agent startup ## Architecture ``` User message → recall (hybrid BM25+vector search) → inject recalled-context block Compaction event → archive messages → extract knowledge facts ``` ### Core Components - **WarmStore** — segment storage with BM25 + vector hybrid search, time-decay, MMR re-ranking - **ColdStore** — JSONL streaming persistence - **KnowledgeStore** — IDF-weighted fact search - **Smart-Trim** — context-event handler: trims old messages, archives, injects recall - **BM25** — full text search with CJK bigram tokenization - **Vector Index** — in-memory HNSW-like vector search with pluggable embeddings (Gemini / hash fallback) - **Compaction Bridge** — archives messages + async knowledge extraction ### Key Features - **Hybrid search** — BM25 + vector cosine with configurable alpha blending - **CJK bigram tokenization** — proper Chinese/Japanese/Korean text handling - **Window recall** — ±2 timeline neighbor expansion for context continuity - **Noise segment filter** — automatically skips heartbeat, no-reply, raw audio, queued messages, and ultra-short content - **Channel prefix stripping** — removes Feishu/WeChat system prefixes before archiving - **Auto embedding model** — defaults to `auto` (Gemini → hash fallback); handles dimension mismatch with background re-embedding - **Non-blocking init** — store loads asynchronously, never blocks agent startup - **Eviction & deduplication** — configurable max-age eviction, cross-compaction dedup - **Redaction** — optional PII stripping ## Tests 11 test files, 90 tests — unit, integration, and E2E (compaction → archive → recall lifecycle, persistence across restart). ## Production Validation Deployed on live Feishu channel. Noise filter removes ~12% junk segments. Gemini embedding (dim=3072) with automatic background re-embedding on model switch. Zero data loss across multiple gateway restarts. ## Breaking Changes None. New plugin — disabled by default, no existing code modified.