dsh-memory-dba 0.1.3 → 0.1.5

package/CHANGELOG.md

@@ -1,5 +1,33 @@
 # Changelog
 
+## 0.1.5 — 2026-10-26
+
+Documentation-only revision. No runtime source file, API, or behavior changed.
+
+- Separated the two state interfaces that the earlier scope text merged into one:
+  **(A) memory-token prefix** (frozen RWKV state → *k* continuous tokens prepended
+  to a frozen LLM's instruction embeddings, trained on downstream task error) and
+  **(B) state-to-text slot decoding** (state → four readable slots → text).
+- Recorded route (A)'s actual standing: 0.5477 vs 0.6774 direct-MLP baseline on 50
+  sealed test episodes, episode-specific per the correct/wrong/zero/random ablation,
+  but token-count ranking unsettled, 64-token seeds incomplete, and behind the MLP
+  baseline on a 192k-sample budget.
+- Bounded every `0/1779` figure to route (B), and named the runtime's actual third
+  path: text summaries plus retrieval.
+
+## 0.1.4 — 2026-10-26
+
+Documentation-only revision. No runtime source file, API, or behavior changed.
+
+- Corrected an inverted research claim. `0.1.2`/`0.1.3` said the v4 state decoder's
+  no-state loss was better than its true-state loss. The recorded run says the
+  opposite: on dev (1,779 cases) true-state token loss was **1.185**, against
+  **1.575** with no state and **1.260** with a mismatched state. The decoder fails
+  on joint multi-label extraction (all-slot exact 0/1779, content-slot accuracy
+  1.6%), not because the frozen state carries no signal; per-label linear probes
+  on the same pooled features reach AUC 0.98.
+- `README.zh-CN.md` now carries the same corrected numbers.
+
 ## 0.1.3 — 2026-10-26
 
 Documentation-only revision. No runtime source file, API, or behavior changed.
@@ -29,8 +57,9 @@ Documentation-only revision. No runtime source file, API, or behavior changed.
   failures without retry; and the default memory directory differing from the old
   Python DBA, which requires an explicit `home` to reuse an existing library.
 - Corrected the research status: state-decoder training on frozen RWKV features was performed (v4, 5,000 steps)
-  but did not meet acceptance (dev all-slot exact 0/1779, no-state loss better than
-  true-state, Not qualified, no final test). It is not described as untrained.
+  but did not meet acceptance (dev all-slot exact 0/1779, Not qualified, no final
+  test). It is not described as untrained. This entry originally also claimed the
+  no-state loss beat the true-state loss; that was wrong and is superseded by 0.1.4.
 
 ## 0.1.1 — 2026-10-26
 

package/README.md

@@ -44,9 +44,16 @@ These are real capabilities of the original Python system, not hypothetical futu
 
 ### 3. Research not yet accepted for use in the overall DBA project
 
+Two state interfaces were tried, and neither validates the other:
+
+- **(A) Memory-token prefix.** A frozen RWKV-7 state passes through an adapter into *k* continuous tokens at the LLM hidden width, which are concatenated in front of the instruction embeddings of a frozen Qwen2.5. Only the adapter and the output head train, on downstream task error — no text is produced in between. This is the interface the architecture-judgment experiment tested: best 0.5477 against a direct-MLP baseline of 0.6774 on 50 sealed test episodes, and the ablation (correct vs wrong-episode vs zero vs random tokens) shows the tokens carry episode-specific information. It is not yet accepted either: the rank among token counts is unsettled, the 64-token seeds are incomplete, and on a small 192k-sample budget memory tokens still trail the MLP baseline — they overtake only after enough training data.
+- **(B) State-to-text slot decoding.** A decoder turns the state into four readable slots and the text is fed onward. All six rounds failed acceptance (best 9/1773 all-slot exact). **Every `0/1779` figure on this page belongs to (B) only and says nothing about (A).**
+
+What the DBA runtime actually uses today is a third path: text summaries plus retrieval. It is neither (A) nor (B).
+
 | Research item | Actual state |
 |---|---|
-| Decoding RWKV state into reliable slots or memory codes | State-decoder training experiments were performed on frozen RWKV features, but they did not reach the acceptance standard. The v4 decoder run completed 5,000 steps; on dev, all-slot exact matches were 0/1779, and the final no-state loss was better than the true-state loss. It was judged **Not qualified** and never entered the final test. There is no usable state-to-slot decoder today. |
+| Decoding RWKV state into readable text slots (route B) | State-decoder training experiments were performed on frozen RWKV features, but they did not reach the acceptance standard. The v4 decoder run completed 5,000 steps; on dev (1,779 cases) all-slot exact matches were **0/1779** and content-slot accuracy was 1.6%. The state itself was not useless: true-state token loss 1.185 beat no-state 1.575 and wrong-state 1.260, and per-label linear probes on the same pooled features reached AUC 0.98 — the information is present, the joint multi-label decode is what fails. It was judged **Not qualified** and never entered the final test. There is no usable state-to-text decoder today. |
 | Cross-model consumption of state | Cloud models consume text summaries and retrieved text, not RWKV state tensors or dedicated memory codes. The interface is not built. |
 | A genuinely resident service | RWKV runs through a local runner/bridge. Continuity across turns is not the same as an accepted always-on service; no resident deployment has been accepted. |
 
@@ -70,7 +77,7 @@ Read these before relying on the plugin as shared or historical memory.
 
 ## Release notes
 
-`0.1.2` and `0.1.3` are documentation-only revisions. No runtime source file or behavior changed in either release; the changes are the scope tables, the limits above, and the corrected release history, so that the npm page matches what is actually delivered. See `CHANGELOG.md` for the full history.
+`0.1.2` through `0.1.5` are documentation-only revisions. No runtime source file or behavior changed in either release; the changes are the scope tables, the limits above, and the corrected release history, so that the npm page matches what is actually delivered. See `CHANGELOG.md` for the full history.
 
 ## Install
 

package/README.zh-CN.md

@@ -1,10 +1,10 @@
 # dsh-memory-dba：功能范围
 
-[English README](README.md) · 版本 0.1.3
+[English README](README.md) · 版本 0.1.5
 
 本包是 DBA（Dual-Brain Architecture，双脑架构）的**文件记忆与检索插件**，不是完整双脑运行时。安装它不会同时安装 RWKV 小脑、连续状态后端或原 Python DBA 的可靠事件处理链。
 
-0.1.2 / 0.1.3 是说明文档修订版：明确交付范围和限制、修正发布历史，没有新增运行时功能，也没有修复下列尚未实现的能力。
+0.1.2–0.1.5 是说明文档修订版：明确交付范围和限制、修正发布历史与一处写反的研究结论，没有新增运行时功能，也没有修复下列尚未实现的能力。
 
 ## 本 npm 包已实现
 
@@ -32,7 +32,14 @@
 
 ## 整体 DBA 仍在研究、未达验收标准
 
-- **RWKV 状态可靠解码为槽位或记忆码**：做过基于冻结 RWKV 特征的解码器训练实验。v4 完成 5000 步，但验证集全槽精确命中为 0/1779，未通过验收，也未进入最终测试；这不等于 RWKV 主干微调已成功或已有可用状态接口。
+试过两种状态接口，两者互不构成证据：
+
+- **（A）记忆 Token 前缀**：冻结的 RWKV-7 状态经适配器变成 k 个连续 token（宽度等于 LLM 隐层），直接拼在冻结 Qwen2.5 的指令嵌入前面；只训练适配器和输出头，用下游任务误差反传，**中间不产生任何文本**。这是架构判定实验验证的路线：50 个密封测试 episode 上最佳 0.5477，对直接 MLP 基线 0.6774；消融（正确 / 错误 episode / 置零 / 随机 token）证明 token 携带 episode 特异信息。它也还没验收：档位名次未定、64 Token 种子不全，而且小预算（192k 样本）下记忆 Token 仍落后 MLP，训够样本才反超。
+- **（B）状态解码为文本槽**：把状态解成四个可读槽位再往下喂。六轮全部未达验收（最好 9/1773 全槽全对）。**本页所有 0/1779 只属于（B），不能用来评价（A）。**
+
+DBA 运行时目前实际走的是第三条：文本摘要 + 检索，既不是（A）也不是（B）。
+
+- **RWKV 状态可靠解码为文本槽（路线 B）**：做过基于冻结 RWKV 特征的解码器训练实验。v4 完成 5000 步，验证集 1779 例里全槽精确命中 0/1779、内容槽准确率 1.6%，未通过验收，也未进入最终测试。但状态本身不是没用：真实状态 token loss 1.185 优于无状态 1.575、错配状态 1.260，同一池化特征上的逐标签线性探针 AUC 达 0.98 —— 信息在，失败的是联合多标签解码。这不等于 RWKV 主干微调已成功或已有可用状态接口。
 - **跨模型状态消费**：云端目前消费文本摘要与检索原文，不是直接接收 RWKV 状态张量或专用记忆码。
 - **经验证的模型常驻服务**：跨轮状态连续与进程/权重始终常驻是两回事，原系统也没有完成常驻部署验收。
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-memory-dba",
-  "version": "0.1.3",
+  "version": "0.1.5",
   "description": "File-based, git-portable long-term memory for DeepSeek Harness: core/stream/workspace layers, BM25 retrieval with CJK bigrams, two-section prompt injection",
   "license": "MIT",
   "repository": {