compound-memory 0.1.0__tar.gz → 0.3.0__tar.gz

{compound_memory-0.1.0/src/compound_memory.egg-info → compound_memory-0.3.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: compound-memory
-Version: 0.1.0
+Version: 0.3.0
 Summary: Local multi-agent shared memory with compounding (MCP server + CLI)
 Author: chinwe
 License-Expression: MIT
@@ -147,4 +147,4 @@ PyPI 版本不可重传，tag 必须与 `pyproject.toml` 的 `version` 一致（
 
 1. **一次性配置**（PyPI → 项目 → Publishing）：owner `chinwe`、repo `compound-memory`、workflow `release.yml`、environment `pypi`。首次发布时项目尚不存在，在 pypi.org 用"pending publisher"预注册即可。
 2. **发布**：`git tag v0.1.0 && git push origin v0.1.0` → `release.yml` 自动 build + `uv publish`。
-3. 发布后 `uvx compound-memory-server` 即为通用安装形态（MCP 配置里的 `command` 也可换成 `uvx`，不再依赖仓库克隆路径）。
+3. 发布后 `uvx --from compound-memory compound-memory-server` 即为通用安装形态（`uvx` 的参数是包名，script 名不同须用 `--from`；MCP 配置里的 `command` 换成 uvx 后不再依赖仓库克隆路径）。

{compound_memory-0.1.0 → compound_memory-0.3.0}/README.md

@@ -118,4 +118,4 @@ PyPI 版本不可重传，tag 必须与 `pyproject.toml` 的 `version` 一致（
 
 1. **一次性配置**（PyPI → 项目 → Publishing）：owner `chinwe`、repo `compound-memory`、workflow `release.yml`、environment `pypi`。首次发布时项目尚不存在，在 pypi.org 用"pending publisher"预注册即可。
 2. **发布**：`git tag v0.1.0 && git push origin v0.1.0` → `release.yml` 自动 build + `uv publish`。
-3. 发布后 `uvx compound-memory-server` 即为通用安装形态（MCP 配置里的 `command` 也可换成 `uvx`，不再依赖仓库克隆路径）。
+3. 发布后 `uvx --from compound-memory compound-memory-server` 即为通用安装形态（`uvx` 的参数是包名，script 名不同须用 `--from`；MCP 配置里的 `command` 换成 uvx 后不再依赖仓库克隆路径）。

{compound_memory-0.1.0 → compound_memory-0.3.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "compound-memory"
-version = "0.1.0"
+version = "0.3.0"
 description = "Local multi-agent shared memory with compounding (MCP server + CLI)"
 readme = "README.md"
 # PEP 639 SPDX 表达式；与 License:: classifier 互斥，故后者已移除

{compound_memory-0.1.0 → compound_memory-0.3.0}/src/compound_memory/__init__.py

@@ -1,3 +1,3 @@
 """compound-memory: local multi-agent shared memory with compounding."""
 
-__version__ = "0.1.0"
+__version__ = "0.2.0"

{compound_memory-0.1.0 → compound_memory-0.3.0}/src/compound_memory/cli.py

@@ -5,12 +5,14 @@ from __future__ import annotations
 import argparse
 import datetime as dt
 import json
+import os
 import sys
 from pathlib import Path
 from typing import Any
 
 from . import __version__
 from .embedding import auto_encoder
+from .extraction import extract, extract_dir
 from .storage import DISTILL_DUP_SIM_THRESHOLD, MEMORY_TYPES, MemoryStore, PROMOTION_USES_THRESHOLD, default_root
 
 
@@ -19,8 +21,13 @@ def _emit(payload: Any) -> None:
 
 
 def _open_store(args: argparse.Namespace) -> MemoryStore:
-    # CLI 与 server 同权：vec extra + 模型就绪即启用向量路，否则自动降级纯词面
-    return MemoryStore(Path(args.root), embedder=auto_encoder())
+    # CLI 与 server 同权：vec extra + 模型就绪即启用向量路，否则自动降级纯词面；
+    # 环境注入的进程身份（若有）同样生效——未设置时保持自报身份模式
+    return MemoryStore(
+        Path(args.root),
+        embedder=auto_encoder(),
+        agent_id=os.environ.get("COMPOUND_MEMORY_AGENT_ID") or None,
+    )
 
 
 def cmd_init(args: argparse.Namespace) -> None:
@@ -29,19 +36,33 @@ def cmd_init(args: argparse.Namespace) -> None:
 
 
 def cmd_write(args: argparse.Namespace) -> None:
-    _emit(_open_store(args).write(content=args.content, type=args.type, source=args.source, ns=args.ns, key=args.key))
+    _emit(
+        _open_store(args).write(
+            content=args.content,
+            type=args.type,
+            source=args.source,
+            ns=args.ns,
+            key=args.key,
+            valid_from=args.valid_from,
+            valid_until=args.valid_until,
+        )
+    )
 
 
 def cmd_search(args: argparse.Namespace) -> None:
     _emit(
         _open_store(args).search(
-            query=args.query, ns=args.ns, top_k=args.top_k, include_neighbors=args.include_neighbors
+            query=args.query,
+            ns=args.ns,
+            top_k=args.top_k,
+            include_neighbors=args.include_neighbors,
+            reader=args.reader,
         )
     )
 
 
 def cmd_get(args: argparse.Namespace) -> None:
-    _emit(_open_store(args).get(args.id))
+    _emit(_open_store(args).get(args.id, reader=args.reader))
 
 
 def cmd_link(args: argparse.Namespace) -> None:
@@ -62,7 +83,7 @@ def cmd_decay(args: argparse.Namespace) -> None:
 
 
 def cmd_revive(args: argparse.Namespace) -> None:
-    _emit(_open_store(args).revive(args.id))
+    _emit(_open_store(args).revive(args.id, reader=args.reader))
 
 
 def cmd_stats(args: argparse.Namespace) -> None:
@@ -88,6 +109,7 @@ def cmd_distill_plan(args: argparse.Namespace) -> None:
             min_uses=args.min_uses,
             min_confidence=args.min_confidence,
             ns=args.ns,
+            reader=args.reader,
         )
     )
 
@@ -106,6 +128,14 @@ def cmd_distill_apply(args: argparse.Namespace) -> None:
     )
 
 
+def cmd_extract(args: argparse.Namespace) -> None:
+    target = Path(args.transcript)
+    if target.is_dir():
+        _emit(extract_dir(target, _open_store(args)))
+    else:
+        _emit(extract(target, _open_store(args)))
+
+
 def cmd_git_log(args: argparse.Namespace) -> None:
     _emit(_open_store(args).git_log(limit=args.limit))
 
@@ -121,22 +151,35 @@ def build_parser() -> argparse.ArgumentParser:
     p = sub.add_parser("write")
     p.add_argument("content"); p.add_argument("type", choices=MEMORY_TYPES)
     p.add_argument("source"); p.add_argument("--ns", default="_shared"); p.add_argument("--key", default=None)
+    p.add_argument("--valid-from", default=None, help="ISO date: fact valid from (annotation)")
+    p.add_argument("--valid-until", default=None,
+                   help="ISO date: fact expires after this day (excluded from search, still readable via get)")
     p.set_defaults(func=cmd_write)
 
     p = sub.add_parser("search")
-    p.add_argument("query"); p.add_argument("--ns", default="_shared"); p.add_argument("--top-k", type=int, default=5)
+    p.add_argument("query")
+    p.add_argument("--ns", default=None,
+                   help="scope: '_shared', 'agent-<name>', or omit for _shared + your own private ns")
+    p.add_argument("--top-k", type=int, default=5)
+    p.add_argument("--reader", default=None, help="caller identity, required for private agent-* namespaces")
     p.add_argument("--no-neighbors", dest="include_neighbors", action="store_false",
                    help="omit embedded one-hop neighbors from hits")
     p.set_defaults(func=cmd_search)
 
-    p = sub.add_parser("get"); p.add_argument("id"); p.set_defaults(func=cmd_get)
+    p = sub.add_parser("get")
+    p.add_argument("id")
+    p.add_argument("--reader", default=None, help="caller identity, required for private agent-* namespaces")
+    p.set_defaults(func=cmd_get)
     p = sub.add_parser("link"); p.add_argument("a"); p.add_argument("b"); p.set_defaults(func=cmd_link)
     p = sub.add_parser("feedback"); p.add_argument("id"); p.add_argument("agent"); p.set_defaults(func=cmd_feedback)
 
     p = sub.add_parser("decay"); p.add_argument("--now", default=None, help="ISO date override (testing)")
     p.set_defaults(func=cmd_decay)
 
-    p = sub.add_parser("revive"); p.add_argument("id"); p.set_defaults(func=cmd_revive)
+    p = sub.add_parser("revive")
+    p.add_argument("id")
+    p.add_argument("--reader", default=None, help="caller identity, required for private agent-* namespaces")
+    p.set_defaults(func=cmd_revive)
     p = sub.add_parser(
         "distill-plan",
         help="scan distillation candidates and print a signal-annotated list",
@@ -148,6 +191,7 @@ def build_parser() -> argparse.ArgumentParser:
     p.add_argument("--min-uses", type=int, default=1, help="activity gate: uses >= this")
     p.add_argument("--min-confidence", type=float, default=0.5, help="activity gate: confidence >= this")
     p.add_argument("--ns", default="_shared")
+    p.add_argument("--reader", default=None, help="caller identity, required for private agent-* namespaces")
     p.set_defaults(func=cmd_distill_plan)
     p = sub.add_parser("distill-apply")
     p.add_argument("content"); p.add_argument("type", choices=MEMORY_TYPES); p.add_argument("source")
@@ -170,6 +214,28 @@ def build_parser() -> argparse.ArgumentParser:
     p.add_argument("--all", action="store_true", help="clear the whole queue")
     p.set_defaults(func=cmd_review_resolve)
     p = sub.add_parser("git-log"); p.add_argument("--limit", type=int, default=5); p.set_defaults(func=cmd_git_log)
+    p = sub.add_parser(
+        "extract",
+        help="scan a session transcript for memory candidates (deterministic pass, no LLM)",
+        description="Deterministic candidate extraction from a session transcript. Transcript shapes are "
+        "auto-detected by content (not filename): WorkBuddy session log jsonl, the ZCode session database "
+        "(~/.zcode/cli/db/db.sqlite, full history), Claude Code session log jsonl, and DeepSeek Harness "
+        "zstd-compressed session files. Pass a directory to batch-scan every supported session file under it "
+        "(<project>/<session>.jsonl layouts and <project>/<session>/session.jsonl.zstd; subagents/ skipped — "
+        "their role:user is the team-lead agent's task brief, not the human's own statement). "
+        "ZCode rollout/model-io snapshots and WorkBuddy traces/ are deliberately unsupported — they keep only "
+        "the most recent / first turns, so accepting them would look like a scan while silently dropping most "
+        "of the history. "
+        "Pattern matching only (statement -> fact, pitfall -> insight); the manifest lands in "
+        "<root>/extract/last-candidates.json. Writing stays with the agent: confirm each candidate "
+        "via memory_write (same-key conflicts still enter the review queue).",
+    )
+    p.add_argument(
+        "transcript",
+        help="path to a session log jsonl (WorkBuddy / Claude Code), the ZCode session database (.sqlite), "
+        "a DeepSeek Harness session file (.jsonl.zstd), or a directory of session logs",
+    )
+    p.set_defaults(func=cmd_extract)
     return parser
 
 

{compound_memory-0.1.0 → compound_memory-0.3.0}/src/compound_memory/embedding.py

@@ -22,6 +22,11 @@ from typing import Callable
 MODEL_REPO_ID = os.environ.get("COMPOUND_MEMORY_EMBEDDING_MODEL", "Xenova/bge-small-zh-v1.5")
 EMBED_DIM = int(os.environ.get("COMPOUND_MEMORY_EMBEDDING_DIM", "512"))
 
+# 单次 onnx run 的批量上限：rebuild 把全库一次喂进来会变成单次巨批 run
+# （LongMemEval 960 条长会话实测 >40 分钟无进度），切块让成本线性可控。
+# 对调用方 encode 仍是一次全量调用——这是批量面优化，与增量索引无关。
+ENCODE_CHUNK = 32
+
 
 def _cache_glob(repo_id: str) -> str:
     """HF 缓存目录 glob：repo id 的 "/" 替换为 "--"（如 a/b → models--a--b）。"""
@@ -76,7 +81,11 @@ class BgeEncoder:
         self._tokenizer: "Tokenizer | None" = None
 
     def encode(self, texts: list[str]) -> list[list[float]]:
-        """批量编码；L2 归一化后的 [CLS] 表示（余弦可直接用作相似度）。"""
+        """批量编码；L2 归一化后的 [CLS] 表示（余弦可直接用作相似度）。
+
+        内部按 ENCODE_CHUNK 切块逐次 run：外部仍是一次全量调用、返回顺序
+        与输入一一对应，只是把单次巨批拆成有界小批（issue #16）。
+        """
         assert VEC_AVAILABLE  # 构造已保证；reassure 类型检查
         if self._session is None or self._tokenizer is None:
             self._tokenizer = Tokenizer.from_file(str(self._tokenizer_path))
@@ -84,13 +93,17 @@ class BgeEncoder:
             self._tokenizer.enable_padding()
             self._session = ort.InferenceSession(str(self._onnx_path), providers=["CPUExecutionProvider"])
         encs = self._tokenizer.encode_batch(texts)
-        feed = {
-            "input_ids": np.array([e.ids for e in encs], dtype=np.int64),
-            "attention_mask": np.array([e.attention_mask for e in encs], dtype=np.int64),
-            "token_type_ids": np.array([e.type_ids for e in encs], dtype=np.int64),
-        }
         names = {i.name for i in self._session.get_inputs()}
-        out = self._session.run(None, {k: v for k, v in feed.items() if k in names})[0]
-        cls = out[:, 0, :]
-        normed = cls / np.linalg.norm(cls, axis=1, keepdims=True)
-        return normed.tolist()
+        out_rows: list[list[float]] = []
+        for start in range(0, len(encs), ENCODE_CHUNK):
+            batch = encs[start : start + ENCODE_CHUNK]
+            feed = {
+                "input_ids": np.array([e.ids for e in batch], dtype=np.int64),
+                "attention_mask": np.array([e.attention_mask for e in batch], dtype=np.int64),
+                "token_type_ids": np.array([e.type_ids for e in batch], dtype=np.int64),
+            }
+            out = self._session.run(None, {k: v for k, v in feed.items() if k in names})[0]
+            cls = out[:, 0, :]
+            normed = cls / np.linalg.norm(cls, axis=1, keepdims=True)
+            out_rows.extend(normed.tolist())
+        return out_rows