memocode 0.2.3__tar.gz → 0.3.2__tar.gz

{memocode-0.2.3 → memocode-0.3.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: memocode
-Version: 0.2.3
+Version: 0.3.2
 Summary: Personal AI coding agent with memory, tool execution, and safety controls
 Author: AssassinCHN
 Requires-Python: >=3.11
@@ -18,9 +18,9 @@ Personal AI coding agent CLI with memory, tool execution, and safety controls.
 ## Features
 
 - **Agentic tool-use loop** — Native function calling (OpenAI / Anthropic compatible), stuck detection, auto mode for fully autonomous task execution
-- **Memory system** — Persistent conversation memory with consolidation, forgetting curve, and cross-session injection
+- **Multi-layer memory system** — Cross-session persistence with structured compression, on-demand recall, and forgetting curve
 - **Project management** — Multiple projects with separate memory DBs and working directories
-- **Built-in tools** — Shell execution, file read/write/edit, glob, grep, web fetch
+- **Built-in tools** — Shell execution, file read/write/edit, glob, grep, web fetch, memory search
 - **Safety layer** — Human mode (zone checks + confirmation prompts) / Auto mode (whitelist + work_dir hard boundaries) / Bypass mode (session-only, explicit confirmation required)
 - **Extensible** — Drop Python files into `~/.mcode/tools/` for custom tools
 
@@ -46,7 +46,7 @@ Configure your model in `~/.mcode/agent.json` (created on first run):
       "model": "MiniMax-M2.7",
       "base_url": "https://api.minimaxi.com/v1",
       "api_key_env": "MINIMAX_API_KEY",
-      "context_window": 1000000,
+      "context_window": 65536,
       "extra_body": {"reasoning_split": true}
     }
   }
@@ -93,11 +93,34 @@ mcode --verbose            # Show full tracebacks on errors
 | `/undo` | Undo the last run — restore code and/or rewind conversation |
 | `/rollback` | Restore a specific backed-up file (file only, for audit) |
 | `/rewind [N]` | Rewind conversation to turn N (conversation only) |
-| `/profile` | Show/edit core memory (user profile) |
+| `/memory` | Show core memory (user profile) |
+| `/memory set <key> <value>` | Write a core memory entry |
+| `/memory del <key>` | Delete a core memory entry |
+| `/memory pin <key>` | Pin an entry (stability=999, never forgets) |
 | | |
 | `!<command>` | Run shell command directly (safety-checked) |
 | `@<path>` | Attach file contents to your message |
 
+## Memory System
+
+Mcode maintains four memory layers that persist across sessions:
+
+| Layer | Scope | Contents | Updated |
+|-------|-------|----------|---------|
+| **Core memory** | Global (all projects) | User traits: communication style, autonomy preference | Every compression + session end |
+| **Project memory** | Per-project | Decisions, architecture, progress, conventions | Every compression + session end |
+| **Recent memory** | Per-project | Compressed summaries of past sessions | Session end; grows indefinitely |
+| **Session history** | Per-project | Current session verbatim + older turns compressed | Each turn |
+
+**How recall works:**
+- Recent session summaries are automatically injected into every turn (newest-first, 4k token cap)
+- The LLM calls `memory_search` when it needs context from older sessions — it decides when and what to search
+- Project memory and core memory are always present in the system prefix (prompt-cache eligible)
+
+**Compression:** When the session context reaches 50% of `context_window`, old turns are replaced with a structured summary (topics, decisions, progress, pending, key context). Raw code and config values are excluded — only decisions and rationale are preserved.
+
+**Forgetting:** Core memory entries decay over time (Ebbinghaus curve). Entries not reinforced by the judge gradually fade, preventing stale user traits from persisting indefinitely.
+
 ## Auto Mode
 
 Auto mode runs the agent fully autonomously — no confirmation prompts, hard safety boundaries (whitelist-only shell commands, writes restricted to `work_dir`).
@@ -131,6 +154,7 @@ Verify by running: <test command>
 | `glob` | Find files by pattern (`**/*.py`) |
 | `grep` | Search file content by regex |
 | `web_fetch` | Fetch a URL, returns readable text (HTML stripped) |
+| `memory_search` | Search past session summaries for relevant context |
 
 ## Custom Tools
 
@@ -167,6 +191,13 @@ mcode/
 │   ├── project_manager.py  # Project registry
 │   ├── audit.py            # Audit log
 │   └── chatmem/            # Memory system
+│       ├── context_manager.py   # Session history, compression, injection
+│       ├── compressor.py        # LLM-based structured summarization
+│       └── memory/
+│           ├── core_memory.py   # User traits (global, with forgetting)
+│           ├── recent_memory.py # Cross-session summaries (per-project)
+│           ├── consolidation.py # Periodic pattern extraction → core memory
+│           └── forgetting.py    # Ebbinghaus decay for core memory
 ├── tools/
 │   ├── file.py             # file_read/write/edit, glob, grep
 │   ├── shell.py            # shell_exec

{memocode-0.2.3 → memocode-0.3.2}/README.md

@@ -6,9 +6,9 @@ Personal AI coding agent CLI with memory, tool execution, and safety controls.
 ## Features
 
 - **Agentic tool-use loop** — Native function calling (OpenAI / Anthropic compatible), stuck detection, auto mode for fully autonomous task execution
-- **Memory system** — Persistent conversation memory with consolidation, forgetting curve, and cross-session injection
+- **Multi-layer memory system** — Cross-session persistence with structured compression, on-demand recall, and forgetting curve
 - **Project management** — Multiple projects with separate memory DBs and working directories
-- **Built-in tools** — Shell execution, file read/write/edit, glob, grep, web fetch
+- **Built-in tools** — Shell execution, file read/write/edit, glob, grep, web fetch, memory search
 - **Safety layer** — Human mode (zone checks + confirmation prompts) / Auto mode (whitelist + work_dir hard boundaries) / Bypass mode (session-only, explicit confirmation required)
 - **Extensible** — Drop Python files into `~/.mcode/tools/` for custom tools
 
@@ -34,7 +34,7 @@ Configure your model in `~/.mcode/agent.json` (created on first run):
       "model": "MiniMax-M2.7",
       "base_url": "https://api.minimaxi.com/v1",
       "api_key_env": "MINIMAX_API_KEY",
-      "context_window": 1000000,
+      "context_window": 65536,
       "extra_body": {"reasoning_split": true}
     }
   }
@@ -81,11 +81,34 @@ mcode --verbose            # Show full tracebacks on errors
 | `/undo` | Undo the last run — restore code and/or rewind conversation |
 | `/rollback` | Restore a specific backed-up file (file only, for audit) |
 | `/rewind [N]` | Rewind conversation to turn N (conversation only) |
-| `/profile` | Show/edit core memory (user profile) |
+| `/memory` | Show core memory (user profile) |
+| `/memory set <key> <value>` | Write a core memory entry |
+| `/memory del <key>` | Delete a core memory entry |
+| `/memory pin <key>` | Pin an entry (stability=999, never forgets) |
 | | |
 | `!<command>` | Run shell command directly (safety-checked) |
 | `@<path>` | Attach file contents to your message |
 
+## Memory System
+
+Mcode maintains four memory layers that persist across sessions:
+
+| Layer | Scope | Contents | Updated |
+|-------|-------|----------|---------|
+| **Core memory** | Global (all projects) | User traits: communication style, autonomy preference | Every compression + session end |
+| **Project memory** | Per-project | Decisions, architecture, progress, conventions | Every compression + session end |
+| **Recent memory** | Per-project | Compressed summaries of past sessions | Session end; grows indefinitely |
+| **Session history** | Per-project | Current session verbatim + older turns compressed | Each turn |
+
+**How recall works:**
+- Recent session summaries are automatically injected into every turn (newest-first, 4k token cap)
+- The LLM calls `memory_search` when it needs context from older sessions — it decides when and what to search
+- Project memory and core memory are always present in the system prefix (prompt-cache eligible)
+
+**Compression:** When the session context reaches 50% of `context_window`, old turns are replaced with a structured summary (topics, decisions, progress, pending, key context). Raw code and config values are excluded — only decisions and rationale are preserved.
+
+**Forgetting:** Core memory entries decay over time (Ebbinghaus curve). Entries not reinforced by the judge gradually fade, preventing stale user traits from persisting indefinitely.
+
 ## Auto Mode
 
 Auto mode runs the agent fully autonomously — no confirmation prompts, hard safety boundaries (whitelist-only shell commands, writes restricted to `work_dir`).
@@ -119,6 +142,7 @@ Verify by running: <test command>
 | `glob` | Find files by pattern (`**/*.py`) |
 | `grep` | Search file content by regex |
 | `web_fetch` | Fetch a URL, returns readable text (HTML stripped) |
+| `memory_search` | Search past session summaries for relevant context |
 
 ## Custom Tools
 
@@ -155,6 +179,13 @@ mcode/
 │   ├── project_manager.py  # Project registry
 │   ├── audit.py            # Audit log
 │   └── chatmem/            # Memory system
+│       ├── context_manager.py   # Session history, compression, injection
+│       ├── compressor.py        # LLM-based structured summarization
+│       └── memory/
+│           ├── core_memory.py   # User traits (global, with forgetting)
+│           ├── recent_memory.py # Cross-session summaries (per-project)
+│           ├── consolidation.py # Periodic pattern extraction → core memory
+│           └── forgetting.py    # Ebbinghaus decay for core memory
 ├── tools/
 │   ├── file.py             # file_read/write/edit, glob, grep
 │   ├── shell.py            # shell_exec

{memocode-0.2.3 → memocode-0.3.2}/control/brain.py

@@ -68,6 +68,8 @@ def _tool_hint(name: str, args: dict) -> str:
         return dim(args.get("pattern", ""))
     if name == "web_fetch":
         return dim(args.get("url", ""))
+    if name == "memory_search":
+        return dim(args.get("query", ""))
     return ""
 
 
@@ -93,42 +95,112 @@ def _print_tool_header_simple(name: str, args: dict):
                 print(dim(f"   {extra}"))
 
 
-def _resolve_task_paths(tasks: list, work_dir: str) -> list:
-    """Legacy helper kept for any remaining callers."""
-    result = []
-    for t in tasks:
-        path_args = {
-            "file_read": ["path"], "file_write": ["path"],
-        }.get(t.tool, [])
-        if not path_args:
-            result.append(t)
-            continue
-        args = dict(t.args)
-        for arg_name in path_args:
-            p = args.get(arg_name, "")
-            if (isinstance(p, str) and p
-                    and not os.path.isabs(p)
-                    and not p.startswith("~")
-                    and "task_" not in p):   # skip {{task_N_output}} placeholders
-                args[arg_name] = os.path.join(work_dir, p)
-        result.append(Task(
-            id=t.id, tool=t.tool, args=args,
-            description=t.description, depends_on=t.depends_on,
-            reversible=t.reversible, backup_paths=t.backup_paths,
-        ))
-    return result
+
+
+_PROJECT_MEMORY_JSON = os.path.join(_DATA_DIR, "project_memory.json")
+
+_DEFAULT_PROJECT_MEMORY = {
+    "dimensions": {
+        "architecture": {"label": "Tech stack, modules, structure",             "stability": 40.0},
+        "decisions":    {"label": "Confirmed design decisions and rationale",   "stability": 35.0},
+        "conventions":  {"label": "Naming, interface, format conventions",      "stability": 35.0},
+        "context":      {"label": "Project background, goals, constraints",     "stability": 30.0},
+        "progress":     {"label": "Current phase, pending key tasks",           "stability": 20.0},
+        "completed":    {"label": "Completed tasks and milestones",             "stability": 25.0},
+    },
+}
+
+def _build_project_judge_prompt(dimensions: dict) -> str:
+    dim_lines = "\n".join(f"- {name}: {cfg.label}" for name, cfg in dimensions.items())
+    dim_names = ", ".join(f'"{name}"' for name in dimensions)
+    return f"""You are a project memory editor. Review the conversation and update the project's knowledge base.
+
+Project memory dimensions:
+{dim_lines}
+
+Existing project memory:
+{{existing}}
+
+New conversation:
+{{text}}
+
+Rules:
+- One entry per dimension. Key must equal the dimension name (e.g. key="decisions").
+- For each dimension, choose one action:
+  SKIP       — no new project-relevant information for this dimension.
+  SUPPLEMENT — new info extends existing (merge into one updated value).
+  UPDATE     — existing info is clearly superseded or corrected.
+  DELETE     — the entire dimension entry is explicitly invalidated with nothing to replace it
+               (e.g. "we're dropping the X feature entirely", "forget everything about the old auth system").
+               Use DELETE only when UPDATE with a replacement value is not possible.
+- Special rules by dimension:
+  completed: always SUPPLEMENT (append new completions, never erase old ones). Never DELETE.
+  progress:  UPDATE freely when the current phase or pending tasks change.
+             ONLY record what the user explicitly stated. Do NOT infer, extrapolate,
+             or add typical/expected follow-up tasks not mentioned in the conversation.
+  decisions: SUPPLEMENT by default; UPDATE only if the user makes an explicit statement reversing a prior decision
+             (e.g. "we're switching from X to Y", "decided to drop X"). Do NOT update based on the assistant
+             discussing alternatives, or the user asking hypothetical questions about other options.
+  architecture/conventions: SUPPLEMENT by default; UPDATE on explicit refactor or migration.
+- Return ONLY entries that changed. Omit SKIPped ones.
+- Exclude: user personal preferences, one-time questions, reasoning processes.
+- Never fabricate or infer information not present in the conversation.
+
+Return a JSON array (empty if nothing changed):
+[{{"key": "<dimension_name>", "value": "<merged or new value>", "dimension": "<dimension_name>", "action": "supplement|update|delete"}}]
+Dimension name must be one of: {dim_names}
+
+Output JSON only:"""
+
+
+def _build_project_judge_prompt_zh(dimensions: dict) -> str:
+    dim_lines = "\n".join(f"- {name}: {cfg.label}" for name, cfg in dimensions.items())
+    dim_names = ", ".join(f'"{name}"' for name in dimensions)
+    return f"""你是一个项目记忆编辑器。审查对话内容，更新项目知识库。
+
+项目记忆维度：
+{dim_lines}
+
+当前项目记忆：
+{{existing}}
+
+新对话：
+{{text}}
+
+规则：
+- 每个维度只有一条记录，key 等于维度名（如 key="decisions"）。
+- 对每个维度，选择以下操作之一：
+  跳过 — 该维度没有新的项目相关信息。
+  补充 — 新信息扩展了已有内容（合并为一个更新后的 value）。
+  修正 — 已有信息被明确推翻或替换。
+  删除 — 整个维度条目被明确作废且无替代内容
+         （如"我们整个X功能都不做了"、"忘掉旧认证系统的一切"）。
+         仅在无法用修正+新值替代时才使用删除。
+- 各维度特殊规则：
+  completed：始终补充（追加新完成内容，不删除旧记录）。不可删除。
+  progress：当前阶段或待办任务变化时，直接修正。
+            只记录用户明确说出的内容，严禁推断、外推或添加"典型后续任务"。
+            用户未提及的待办事项一律不写。
+  decisions：默认补充；仅当用户明确陈述推翻先前决策时才修正
+             （如"我们从X换成Y了"、"决定放弃X"）。不得因为
+             AI 讨论替代方案、或用户询问假设性问题而修正。
+  architecture/conventions：默认补充；仅在明确重构或迁移时修正。
+- 只返回发生变化的条目，跳过的不返回。
+- 不包括：用户个人偏好、一次性问题、推理过程。
+- 严禁捏造或推断对话中未出现的信息。
+
+返回JSON数组（无变化则返回空数组）：
+[{{"key": "<维度名>", "value": "<合并后或新的内容>", "dimension": "<维度名>", "action": "supplement|update|delete"}}]
+维度名必须是以下之一：{dim_names}
+
+只输出JSON："""
 
 
 _DEFAULT_CHATMEM = {
     "llm": {},
     "dimensions": {
-        "identity":      {"label": "Who I am",         "stability": 50.0},
-        "values":        {"label": "What I believe",   "stability": 40.0},
-        "goals":         {"label": "What I want",      "stability": 30.0},
-        "preferences":   {"label": "Likes / dislikes", "stability": 25.0},
-        "capabilities":  {"label": "What I can do",    "stability": 30.0},
-        "emotional":     {"label": "How I react",      "stability": 20.0},
-        "autobiography": {"label": "Key experiences",  "stability": 35.0},
+        "communication": {"label": "Expression style, language preference, tone, detail level",                                        "stability": 45.0},
+        "autonomy":      {"label": "How much decision authority delegated to LLM: executes directly vs presents options vs validates", "stability": 40.0},
     },
 }
 
@@ -170,11 +242,11 @@ _DEFAULT_AGENT = {
 }
 
 
-def _init_data_dir() -> tuple[str, str]:
+def _init_data_dir() -> tuple[str, str, str]:
     """
-    Ensure ~/.mcode/ exists with agent.json and chatmem.json.
+    Ensure ~/.mcode/ exists with agent.json, chatmem.json, project_memory.json.
     Migrates from control/ if old files exist and new ones don't.
-    Returns (agent_json_path, chatmem_json_path).
+    Returns (agent_json_path, chatmem_json_path, project_memory_json_path).
     """
     import json as _json
     import shutil as _shutil
@@ -184,6 +256,7 @@ def _init_data_dir() -> tuple[str, str]:
 
     agent_path = os.path.join(_DATA_DIR, "agent.json")
     chatmem_path = os.path.join(_DATA_DIR, "chatmem.json")
+    project_memory_path = _PROJECT_MEMORY_JSON
 
     # Migrate agent.json from old location
     old_agent = os.path.join(_CONTROL, "agent.json")
@@ -205,7 +278,13 @@ def _init_data_dir() -> tuple[str, str]:
                 _json.dump(_DEFAULT_CHATMEM, f, indent=2, ensure_ascii=False)
                 f.write("\n")
 
-    return agent_path, chatmem_path
+    # Create project_memory.json with default project dimensions
+    if not os.path.exists(project_memory_path):
+        with open(project_memory_path, "w") as f:
+            _json.dump(_DEFAULT_PROJECT_MEMORY, f, indent=2, ensure_ascii=False)
+            f.write("\n")
+
+    return agent_path, chatmem_path, project_memory_path
 
 
 def _sync_active_model(agent_cfg: dict, chatmem_json_path: str | None):
@@ -227,7 +306,7 @@ def _sync_active_model(agent_cfg: dict, chatmem_json_path: str | None):
                 profile = model_cfg
                 agent_cfg["active_model"] = active
                 # Save the updated active_model back to agent.json
-                agent_path, _ = _init_data_dir()
+                agent_path, *_ = _init_data_dir()
                 with open(agent_path) as f:
                     import json as _json
                     data = _json.load(f)
@@ -260,7 +339,7 @@ class Brain:
         no_lock: bool = False,
     ):
         # Ensure ~/.mcode/ exists; migrate old files if needed
-        agent_path, chatmem_path = _init_data_dir()
+        agent_path, chatmem_path, project_memory_path = _init_data_dir()
 
         self._agent_cfg_path = agent_path
         self._config_path = config_path or chatmem_path
@@ -320,6 +399,11 @@ class Brain:
         else:
             self.projects.find_or_create_for_cwd()
 
+        # Project memory — per-project CoreMemory in the same project DB
+        self.project_memory = self._make_project_memory(
+            cfg, _api_key, project_memory_path
+        )
+
         # Memory layer — one DB per project, core DB shared
         self.memory = ContextManager.create(
             config=cfg,
@@ -327,6 +411,13 @@ class Brain:
             db_path=self.projects.project_db_path,
             core_db_path=self.projects.core_db_path,
             no_lock=no_lock,
+            extra_context_fn=(
+                self.project_memory.to_context_string if self.project_memory else None
+            ),
+            on_compress_fn=(
+                self._run_project_judge if self.project_memory else None
+            ),
+            auto_recall=False,
         )
 
         # Audit log
@@ -341,6 +432,7 @@ class Brain:
         self.registry.register(GLOB_TOOL)
         self.registry.register(GREP_TOOL)
         self.registry.register(WEB_FETCH_TOOL)
+        self.registry.register(self._make_memory_search_tool())
         load_external_tools(self.registry)
 
         self._max_tool_iter: int = agent_cfg.get("max_tool_iter", 40)
@@ -406,7 +498,8 @@ class Brain:
                 "If the user asks you to fetch a URL or HTTP endpoint, call web_fetch — never fabricate the response. "
                 "Before editing a file, always read it first with file_read, then proceed to make the edit without stopping. "
                 "For targeted changes use file_edit; for new files use file_write. "
-                "If a task requires multiple tool calls, keep calling tools until the task is fully done — do not report completion until all actions have been executed via tools."
+                "If a task requires multiple tool calls, keep calling tools until the task is fully done — do not report completion until all actions have been executed via tools. "
+                "If the user asks about something that may have been discussed in a previous session and you don't have that information in the current context, call memory_search before concluding it is unknown or undecided."
             )
         user_msg = [
             {"role": "system", "content": _system},
@@ -627,7 +720,8 @@ class Brain:
             return True
 
         if self.auto_mode:
-            whitelist = self._agent_cfg.get("auto_whitelist", None)
+            user_wl = self._agent_cfg.get("auto_whitelist", None)
+            whitelist = list(set(_safety.DEFAULT_AUTO_WHITELIST) | set(user_wl)) if user_wl else None
             result = _safety.check_auto(task, work_dir=work_dir, whitelist=whitelist)
             if result.violation:
                 print(red(f"\n🛑  AUTO-BLOCKED  {result.reason}"))
@@ -789,20 +883,115 @@ class Brain:
     # ------------------------------------------------------------------
 
     def switch_project(self, name: str):
-        self.memory.end_session()
+        self.end_session()
         self.projects.switch_project(name)
         self._reinit_memory()
 
+    def _make_memory_search_tool(self):
+        """Build a memory_search Tool that searches recent_memory for relevant past context."""
+        from tools.registry import Tool, ToolSchema
+        recent_memory = self.memory.recent_memory
+
+        def _search(query: str, top_k: int = 5) -> str:
+            results = recent_memory.search(query, top_k=top_k, min_score=0.0)
+            if not results:
+                return "No relevant memory found."
+            return "\n\n---\n\n".join(results)
+
+        return Tool(
+            schema=ToolSchema(
+                name="memory_search",
+                description=(
+                    "Search compressed summaries of past conversation sessions for relevant context. "
+                    "ALWAYS call this before saying something is unknown, undecided, or not recorded — "
+                    "the information may exist in a previous session even if it is not in the current context. "
+                    "Use key terms from the user's question as the query.\n"
+                    "搜索过去对话的压缩摘要，获取相关上下文。"
+                    "在说某事未知、未决定或未记录之前，必须先调用此工具——"
+                    "即使当前上下文中没有该信息，它也可能存在于之前的会话中。"
+                    "用用户问题中的关键词作为查询词。"
+                ),
+                parameters={
+                    "type": "object",
+                    "properties": {
+                        "query": {
+                            "type": "string",
+                            "description": "Key terms to search for (e.g. '8080 端口', 'authentication design')",
+                        },
+                        "top_k": {
+                            "type": "integer",
+                            "description": "Max results to return (default 5)",
+                            "default": 5,
+                        },
+                    },
+                    "required": ["query"],
+                },
+            ),
+            fn=_search,
+        )
+
+    def _make_project_memory(self, cfg: "ContextConfig", api_key: str, project_memory_path: str):
+        """Instantiate a CoreMemory for project-scoped knowledge."""
+        import json as _json
+        from chatmem.memory.core_memory import CoreMemory
+        from chatmem.config import DimensionConfig
+
+        try:
+            with open(project_memory_path) as f:
+                pm_data = _json.load(f)
+            dims = {
+                k: DimensionConfig(label=v["label"], stability=float(v["stability"]))
+                for k, v in pm_data.get("dimensions", {}).items()
+            }
+        except Exception:
+            dims = {}
+        if not dims:
+            return None
+
+        compress_model = cfg.llm.compress_model or cfg.llm.model
+        _THINKING_KEYS = {"reasoning_split", "thinking", "reasoning_effort"}
+        extra_body = (
+            {k: v for k, v in cfg.llm.extra_body.items() if k not in _THINKING_KEYS}
+            if cfg.llm.extra_body else None
+        ) or None
+        provider = getattr(cfg.llm, "provider", "openai") or "openai"
+
+        pm = CoreMemory(
+            db_path=self.projects.project_db_path,
+            api_key=api_key,
+            model=compress_model,
+            base_url=cfg.llm.base_url,
+            dimensions=dims,
+            extra_body=extra_body,
+            provider=provider,
+            context_label="Project Memory",
+        )
+
+        # Override judge prompts with project-specific framing
+        pm._judge_prompt_en = _build_project_judge_prompt(dims)
+        pm._judge_prompt_zh = _build_project_judge_prompt_zh(dims)
+        return pm
+
     def _reinit_memory(self):
         if getattr(self, "memory", None) is not None:
             self.memory.close()
         cfg = ContextConfig.from_json(self._config_path) if self._config_path else ContextConfig()
+        api_key = cfg.llm.resolve_api_key()
+        self.project_memory = self._make_project_memory(cfg, api_key, _PROJECT_MEMORY_JSON)
         self.memory = ContextManager.create(
             config=cfg,
-            api_key=cfg.llm.resolve_api_key(),
+            api_key=api_key,
             db_path=self.projects.project_db_path,
             core_db_path=self.projects.core_db_path,
+            extra_context_fn=(
+                self.project_memory.to_context_string if self.project_memory else None
+            ),
+            on_compress_fn=(
+                self._run_project_judge if self.project_memory else None
+            ),
+            auto_recall=False,
         )
+        self.registry.register(self._make_memory_search_tool())
 
     def _save_agent_cfg(self):
         """Persist current _agent_cfg to agent.json."""
@@ -815,5 +1004,35 @@ class Brain:
     # Session lifecycle
     # ------------------------------------------------------------------
 
+    def _run_project_judge(self, text: str):
+        if not self.project_memory:
+            return
+        import logging as _logging
+        logger = _logging.getLogger("chatmem")
+        try:
+            written = self.project_memory.judge_and_update(text)
+            if written:
+                for e in written:
+                    if e.get("action") == "delete":
+                        logger.info("[project] judge deleted  key=%r", e["key"])
+                    else:
+                        logger.info("[project] judge wrote  key=%r  dim=%s  value=%r",
+                                    e["key"], e.get("dimension"), e["value"][:60])
+            else:
+                logger.info("[project] judge returned empty")
+        except Exception as exc:
+            logger.warning("[project] judge error: %s", exc)
+
     def end_session(self):
+        # Capture verbatim history before end_session clears it
+        verbatim = [m for m in self.memory._history if not m.get("_is_summary")]
+        conv_text = "\n".join(
+            f"{m['role']}: {m.get('content', '')}" for m in verbatim
+        ) if verbatim else ""
+
         self.memory.end_session()
+
+        if conv_text:
+            threading.Thread(
+                target=self._run_project_judge, args=(conv_text,), daemon=True
+            ).start()