memorix 1.9.0 → 1.9.1

package/CHANGELOG.md

@@ -4,6 +4,34 @@ All notable changes to this project will be documented in this file.
 
 ## [Unreleased]
 
+## [1.9.1] - 2026-09-08
+
+### Changed
+- **Large-scale retrieval** -- SQLite FTS5 is now the persistent lexical
+  candidate index. It stays synchronized on insert, update, delete, and can be
+  rebuilt from the durable observations table.
+- **Disk-backed semantic retrieval** -- the optional local LanceDB shadow index
+  stores only vector IDs and metadata, uses HNSW with scalar quantization once
+  the corpus is large enough, and fuses bounded semantic candidates with FTS5
+  results. Search falls back to Orama when the optional native package is not
+  available.
+- **Lazy large-corpus startup** -- projects above the Orama hydration threshold
+  keep SQLite as the source of truth and do not load the complete observation
+  corpus into the interactive agent process. Detail, export, and maintenance
+  paths can still load the data they explicitly need.
+- **Large-store acceptance** -- adds regression coverage and verified gates for
+  1k, 10k, 40k, and 100k records without reducing the durable memory corpus.
+- **Post-threshold vector writes** -- newly generated vectors still reach the
+  persistent semantic index after the in-memory Orama corpus is released.
+
+### Fixed
+- **Stale lexical index on upsert** -- replace-style SQLite writes could leave
+  old terms in an external-content FTS index. Observation upserts now preserve
+  trigger-visible row updates so old terms are removed.
+- **Optional native dependency bundling** -- LanceDB is resolved at runtime so
+  unsupported platforms or installations without its native binary continue
+  to build and use the FTS5/Orama fallback.
+
 ## [1.9.0] - 2026-09-07
 
 ### Added

package/README.md

@@ -63,7 +63,11 @@ Use Claude Code today, Codex tomorrow, and Cursor in the afternoon. The agent ca
 | Static rule files drift | Gotchas, fixes, and project skills evolve from real work |
 | Parallel agent work gets messy | `memorix orchestrate` coordinates task context, handoffs, locks, verification, and review loops |
 
-Memorix is local-first. SQLite is the canonical store, Orama handles search, and LLM-backed formation/embedding is optional. Without model keys, Memorix still works with local full-text retrieval.
+Memorix is local-first. SQLite is the canonical store. Small projects use the
+in-process Orama path; larger projects use a persistent SQLite FTS5 candidate
+index and, when available, an optional local LanceDB semantic shadow index.
+Both indexes are rebuildable and are never a limit on how many durable memories
+you can keep. LLM-backed formation and embedding remain optional.
 
 ### Capabilities
 

package/README.zh-CN.md

@@ -62,7 +62,7 @@ Memorix 给你已经在用的 AI 编程 Agent 加上一套共享、可检索的
 | 静态规则文件容易过期 | 坑点、修复和项目技能从真实工作中持续沉淀 |
 | 并行 Agent 工作容易乱 | `memorix orchestrate` 负责协调任务上下文、交接、文件锁、验证和 review 流程 |
 
-Memorix 是本地优先的。SQLite 是权威存储，Orama 负责搜索，LLM 记忆整理和 embedding 是可选能力。没有模型 key 时，Memorix 仍然可以用本地全文检索工作。
+Memorix 是本地优先的。SQLite 是权威存储。小项目使用进程内 Orama；数据量较大时，关键词检索使用持久化的 SQLite FTS5 候选索引，并在可用时使用本地 LanceDB 语义影子索引。索引都可以重建，也不会限制你能长期保存多少条记忆。LLM 记忆整理和 embedding 仍然是可选能力。
 
 ### 能力矩阵