Hermes LCM Context Management

Hermes LCM 上下文管理

Skill by ara.so — Hermes Skills collection.

由 ara.so 开发的技能 — Hermes 技能集合。

Overview

概述

Hermes-LCM is a lossless context management plugin for Hermes Agent that prevents message loss during context compression. Instead of replacing old messages with flat summaries, it:

Stores all messages in SQLite before compaction
Compacts old context into a hierarchical summary DAG
Provides agent tools to drill back into compacted material
Maintains source lineage for filtered retrieval
Externalizes large payloads to prevent bloat

Key difference from built-in compression: LCM makes recall part of the active context engine with drill-down tools (

lcm_grep

,

lcm_expand

,

lcm_expand_query

) rather than relying on auxiliary cross-session search.

Hermes-LCM是一款适用于Hermes Agent的无损上下文管理插件，可防止上下文压缩过程中的消息丢失。它不会用扁平化摘要替换旧消息，而是：

在压缩前将所有消息存储到SQLite中
将旧上下文压缩为分层摘要DAG
为Agent提供工具以回溯压缩后的内容
维护源谱系以支持过滤检索
将大负载外置以避免内容膨胀

与内置压缩的核心区别：LCM通过钻取工具（

lcm_grep

、

lcm_expand

、

lcm_expand_query

）将召回功能整合到主动上下文引擎中，而非依赖辅助的跨会话搜索。

Installation

安装

Standard Installation

标准安装

Clone into Hermes plugins directory:

bash

undefined

克隆到Hermes插件目录：

bash

undefined

General user plugin (all profiles)

通用用户插件（所有配置文件）

git clone https://github.com/stephenschoettler/hermes-lcm
~/.hermes/plugins/hermes-lcm

Profile-specific install

特定配置文件安装

git clone https://github.com/stephenschoettler/hermes-lcm
~/.hermes/profiles/myprofile/plugins/hermes-lcm

undefined

git clone https://github.com/stephenschoettler/hermes-lcm
~/.hermes/profiles/myprofile/plugins/hermes-lcm

undefined

Symlink Installation

符号链接安装

From an existing checkout:

bash

cd hermes-lcm
./scripts/install.sh

从已有的检出目录执行：

bash

cd hermes-lcm
./scripts/install.sh

Profile-specific

特定配置文件

HERMES_PROFILE=myprofile ./scripts/install.sh

undefined

HERMES_PROFILE=myprofile ./scripts/install.sh

undefined

Configuration

配置

Enable in Hermes config (YAML):

yaml

plugins:
  enabled:
    - hermes-lcm

context:
  engine: lcm

在Hermes配置（YAML）中启用：

yaml

plugins:
  enabled:
    - hermes-lcm

context:
  engine: lcm

Keep compression enabled - LCM needs this gate

保持压缩启用 - LCM需要此开关

compression: enabled: true


Restart Hermes after configuration changes.

compression: enabled: true


配置更改后重启Hermes。

Verification

验证

bash

hermes plugins

Expected output includes:

Plugin list shows
```
hermes-lcm
```
Context engine shows
```
lcm
```

Tools include:

lcm_grep

,

lcm_describe

,

lcm_expand

,

lcm_expand_query

,

lcm_status

,

lcm_doctor

,

lcm_load_session

bash

hermes plugins

预期输出应包含：

插件列表显示
```
hermes-lcm
```
上下文引擎显示
```
lcm
```

工具包含：

lcm_grep

,

lcm_describe

,

lcm_expand

,

lcm_expand_query

,

lcm_status

,

lcm_doctor

,

lcm_load_session

Core Concepts

核心概念

Message Storage

消息存储

LCM stores messages in SQLite before compaction happens:

~/.hermes/profiles/<profile>/lcm.db  # Default path

LCM在压缩前将消息存储到SQLite中：

~/.hermes/profiles/<profile>/lcm.db  # 默认路径

Summary DAG

摘要DAG

Old messages are compacted into hierarchical summary nodes:

Raw messages → Leaf summaries → Branch summaries → Root

Each node tracks:

Descendant message count
Source lineage (for filtering)
Depth in DAG
Token counts

旧消息被压缩为分层摘要节点：

原始消息 → 叶子摘要 → 分支摘要 → 根节点

每个节点追踪：

子消息数量
源谱系（用于过滤）
在DAG中的深度
令牌数量

Bounded Recovery

有限恢复

Agent can page back into compacted material without flooding active context:

```
lcm_grep
```
: Search raw messages and summaries
```
lcm_describe
```
: Get summary metadata
```
lcm_expand
```
: Retrieve child summaries or raw messages
```
lcm_expand_query
```
: Synthesize answer from DAG material

Agent可以回溯压缩内容而不会占用过多主动上下文：

```
lcm_grep
```
：搜索原始消息和摘要
```
lcm_describe
```
：获取摘要元数据
```
lcm_expand
```
：检索子摘要或原始消息
```
lcm_expand_query
```
：从DAG内容中合成答案

Environment Configuration

环境配置

Core Settings

核心设置

bash

undefined

bash

undefined

Compaction trigger (fraction of context window)

压缩触发阈值（上下文窗口占比）

export LCM_CONTEXT_THRESHOLD=0.75

Recent messages protected from compaction

受保护不被压缩的最新消息数量

export LCM_FRESH_TAIL_COUNT=64

Raw backlog floor before leaf compaction

叶子压缩前的原始消息最小令牌数

export LCM_LEAF_CHUNK_TOKENS=20000

Enable dynamic chunk-sized leaf compaction

启用动态块大小的叶子压缩

export LCM_DYNAMIC_LEAF_CHUNK_ENABLED=false export LCM_DYNAMIC_LEAF_CHUNK_MAX=40000

undefined

export LCM_DYNAMIC_LEAF_CHUNK_ENABLED=false export LCM_DYNAMIC_LEAF_CHUNK_MAX=40000

undefined

Session Management

会话管理

bash

undefined

bash

undefined

DAG depth retained after /new (-1 all, 0 none)

/new命令后保留的DAG深度（-1表示全部，0表示无）

export LCM_NEW_SESSION_RETAIN_DEPTH=2

Exclude sessions from LCM storage (glob patterns)

排除LCM存储的会话（通配符模式）

export LCM_IGNORE_SESSION_PATTERNS="test-,debug-"

Keep sessions read-only (glob patterns)

设置会话为只读（通配符模式）

export LCM_STATELESS_SESSION_PATTERNS="readonly-*"

Exclude messages by content regex (comma-separated)

按内容正则排除消息（逗号分隔）

export LCM_IGNORE_MESSAGE_PATTERNS="^SYSTEM:,^[INTERNAL]"

undefined

export LCM_IGNORE_MESSAGE_PATTERNS="^SYSTEM:,^[INTERNAL]"

undefined

Large Payload Handling

大负载处理

bash

undefined

bash

undefined

Store oversized payloads externally

将超大负载存储到外部

export LCM_LARGE_OUTPUT_EXTERNALIZATION_ENABLED=true export LCM_LARGE_OUTPUT_EXTERNALIZATION_THRESHOLD_CHARS=12000

Compact already-externalized tool results

压缩已外置的工具结果

export LCM_LARGE_OUTPUT_TRANSCRIPT_GC_ENABLED=false

undefined

export LCM_LARGE_OUTPUT_TRANSCRIPT_GC_ENABLED=false

undefined

Model Overrides

模型覆盖

bash

undefined

bash

undefined

Override summarization model

覆盖摘要模型

export LCM_SUMMARY_MODEL=claude-3-5-sonnet-20241022

Override expansion synthesis model

覆盖扩展合成模型

export LCM_EXPANSION_MODEL=gpt-4-turbo export LCM_EXPANSION_CONTEXT_TOKENS=32000

Timeouts

超时设置

export LCM_SUMMARY_TIMEOUT_MS=60000 export LCM_EXPANSION_TIMEOUT_MS=120000

undefined

export LCM_SUMMARY_TIMEOUT_MS=60000 export LCM_EXPANSION_TIMEOUT_MS=120000

undefined

Advanced

高级设置

bash

undefined

bash

undefined

Custom database path

自定义数据库路径

export LCM_DATABASE_PATH=/custom/path/lcm.db

Enable slash commands

启用斜杠命令

export LCM_ENABLE_SLASH_COMMAND=true

Allow destructive doctor clean operations

允许doctor工具执行破坏性清理操作

export LCM_DOCTOR_CLEAN_APPLY_ENABLED=false

Critical pressure bypass (0.0 = disabled)

临界预算压力绕过（0.0表示禁用）

export LCM_CRITICAL_BUDGET_PRESSURE_RATIO=0.0

undefined

export LCM_CRITICAL_BUDGET_PRESSURE_RATIO=0.0

undefined

Tuning for Long Context Models

长上下文模型调优

For large context windows, tune threshold to avoid excessive prompt costs:

Calculation:

compaction_trigger = effective_context_window * LCM_CONTEXT_THRESHOLD

Examples:

Context Window	Desired Trigger	Threshold	Use Case
128K	96K	0.75	Standard
200K	140K	0.70	Balanced
400K	240K	0.60	Long context
1M	250K	0.25	Cost-optimized
1M	400K	0.40	Balanced large
1M	600K	0.60	Max raw context

Example configuration for 1M token model:

bash

undefined

对于大上下文窗口，调优阈值以避免过高的提示成本：

计算公式:

压缩触发值 = 有效上下文窗口 * LCM_CONTEXT_THRESHOLD

示例:

上下文窗口	期望触发值	阈值	使用场景
128K	96K	0.75	标准场景
200K	140K	0.70	平衡场景
400K	240K	0.60	长上下文场景
1M	250K	0.25	成本优化场景
1M	400K	0.40	平衡大窗口场景
1M	600K	0.60	最大原始上下文场景

1M令牌模型的配置示例:

bash

undefined

Cost-optimized: trigger at 300K tokens

成本优化：在300K令牌时触发

export LCM_CONTEXT_THRESHOLD=0.30

Balanced: trigger at 400K tokens

平衡场景：在400K令牌时触发

export LCM_CONTEXT_THRESHOLD=0.40

Max context: trigger at 600K tokens

最大上下文：在600K令牌时触发

export LCM_CONTEXT_THRESHOLD=0.60

undefined

export LCM_CONTEXT_THRESHOLD=0.60

undefined

Agent Tools Usage

Agent工具使用

lcm_status

Get current LCM state:

python

undefined

获取当前LCM状态：

python

undefined

Agent calls this to check compression state

Agent调用此工具检查压缩状态

{ "tool": "lcm_status", "params": {} }


Returns:
- Session ID
- Threshold tokens
- Current prompt tokens
- Raw message count
- Summary DAG structure
- Storage path
- Git commit (if source checkout)

{ "tool": "lcm_status", "params": {} }


返回内容：
- 会话ID
- 阈值令牌数
- 当前提示令牌数
- 原始消息数量
- 摘要DAG结构
- 存储路径
- Git提交记录（如果是源码检出）

lcm_grep

Search messages and summaries:

python

undefined

搜索消息和摘要：

python

undefined

Search for keyword in raw messages

在原始消息中搜索关键词

{ "tool": "lcm_grep", "params": { "pattern": "database schema", "search_raw": true, "search_summaries": false, "max_results": 10 } }

Search summaries only

仅搜索摘要

{ "tool": "lcm_grep", "params": { "pattern": "migration", "search_raw": false, "search_summaries": true, "max_results": 5 } }

Filter by source (files/tools mentioned)

按来源过滤（提及的文件/工具）

{ "tool": "lcm_grep", "params": { "pattern": "error", "source_filter": "src/database.py", "search_raw": true } }

undefined

{ "tool": "lcm_grep", "params": { "pattern": "error", "source_filter": "src/database.py", "search_raw": true } }

undefined

lcm_describe

Get summary node metadata:

python

undefined

获取摘要节点元数据：

python

undefined

Describe a specific summary node

描述特定摘要节点

{ "tool": "lcm_describe", "params": { "summary_id": "s_abc123" } }


Returns:
- Summary text
- Descendant count
- Token counts
- Depth in DAG
- Source lineage

{ "tool": "lcm_describe", "params": { "summary_id": "s_abc123" } }


返回内容：
- 摘要文本
- 子节点数量
- 令牌数
- 在DAG中的深度
- 源谱系

lcm_expand

Retrieve child summaries or raw messages:

python

undefined

检索子摘要或原始消息：

python

undefined

Expand a summary to see its children

展开摘要查看其子节点

{ "tool": "lcm_expand", "params": { "summary_id": "s_abc123", "max_children": 5, "max_raw": 10 } }

Get raw messages with source filter

按来源过滤获取原始消息

{ "tool": "lcm_expand", "params": { "summary_id": "s_abc123", "source_filter": "config.py", "max_raw": 20 } }

undefined

{ "tool": "lcm_expand", "params": { "summary_id": "s_abc123", "source_filter": "config.py", "max_raw": 20 } }

undefined

lcm_expand_query

Synthesize answer from DAG material using auxiliary LLM:

python

undefined

使用辅助LLM从DAG内容中合成答案：

python

undefined

Ask a question about compacted history

查询压缩历史中的内容

{ "tool": "lcm_expand_query", "params": { "query": "What database migrations were discussed earlier?", "summary_id": "s_abc123", # optional: scope to subtree "max_raw": 50 } }


This tool:
1. Retrieves relevant raw messages and summaries
2. Calls auxiliary LLM with query + material
3. Returns synthesized answer
4. Uses `LCM_EXPANSION_MODEL` and `LCM_EXPANSION_CONTEXT_TOKENS`

{ "tool": "lcm_expand_query", "params": { "query": "之前讨论过哪些数据库迁移？", "summary_id": "s_abc123", # 可选：限定子树范围 "max_raw": 50 } }


此工具执行以下操作：
1. 检索相关原始消息和摘要
2. 使用查询内容+调用辅助LLM
3. 返回合成答案
4. 使用`LCM_EXPANSION_MODEL`和`LCM_EXPANSION_CONTEXT_TOKENS`配置

lcm_load_session

Load historical session into current context:

python

undefined

将历史会话加载到当前上下文：

python

undefined

Load previous session

加载之前的会话

{ "tool": "lcm_load_session", "params": { "session_id": "previous-session-abc123" } }

undefined

{ "tool": "lcm_load_session", "params": { "session_id": "previous-session-abc123" } }

undefined

lcm_doctor

Diagnose and repair LCM state:

python

undefined

诊断并修复LCM状态：

python

undefined

Check for issues

检查问题

{ "tool": "lcm_doctor", "params": { "action": "check" } }

Preview cleanup (dry run)

预览清理（试运行）

{ "tool": "lcm_doctor", "params": { "action": "clean_preview" } }

Apply cleanup (requires LCM_DOCTOR_CLEAN_APPLY_ENABLED=true)

执行清理（需要LCM_DOCTOR_CLEAN_APPLY_ENABLED=true）

{ "tool": "lcm_doctor", "params": { "action": "clean_apply" } }

undefined

{ "tool": "lcm_doctor", "params": { "action": "clean_apply" } }

undefined

Slash Commands (Optional)

斜杠命令（可选）

Enable with

LCM_ENABLE_SLASH_COMMAND=true

:

bash

undefined

设置

LCM_ENABLE_SLASH_COMMAND=true

启用：

bash

undefined

Check status

检查状态

/lcm status

Search

搜索

/lcm grep pattern search_raw=true

Describe summary

描述摘要

/lcm describe summary_id=s_abc123

Expand

展开

/lcm expand summary_id=s_abc123 max_raw=20

Query

查询

/lcm query What was discussed about the API?

Doctor

诊断

/lcm doctor check /lcm doctor clean_preview

undefined

/lcm doctor check /lcm doctor clean_preview

undefined

Common Patterns

常见使用模式

Initial Setup After Installation

安装后的初始设置

bash

undefined

bash

undefined

1. Verify installation

1. 验证安装

hermes plugins

2. Send a test message to initialize session

2. 发送测试消息初始化会话

hermes chat "Hello"

3. Check LCM status

3. 检查LCM状态

hermes chat "Can you run lcm_status?"

4. Verify tools are available

4. 验证工具可用

Agent should have access to lcm_grep, lcm_expand, etc.

Agent应能访问lcm_grep、lcm_expand等工具

undefined

undefined

Recovering Lost Context

恢复丢失的上下文

python

undefined

python

undefined

Scenario: Agent forgot details about earlier discussion

场景：Agent忘记了早期讨论的细节

1. Search for topic

1. 搜索主题

{ "tool": "lcm_grep", "params": { "pattern": "API authentication", "search_raw": true, "search_summaries": true, "max_results": 10 } }

2. Expand relevant summary

2. 展开相关摘要

{ "tool": "lcm_expand", "params": { "summary_id": "s_found_in_grep", "max_raw": 20 } }

3. Synthesize answer

3. 合成答案

{ "tool": "lcm_expand_query", "params": { "query": "What authentication method did we decide to use?", "summary_id": "s_found_in_grep" } }

undefined

{ "tool": "lcm_expand_query", "params": { "query": "我们决定使用哪种认证方式？", "summary_id": "s_found_in_grep" } }

undefined

Debugging Compaction Issues

调试压缩问题

bash

undefined

bash

undefined

Check current state

检查当前状态

export LCM_ENABLE_SLASH_COMMAND=true hermes chat "/lcm status"

Run diagnostics

运行诊断

hermes chat "/lcm doctor check"

Preview cleanup

预览清理

hermes chat "/lcm doctor clean_preview"

Check for orphaned summaries or raw messages

检查孤立摘要或原始消息

Doctor will report:

Doctor会报告：

- Orphaned summaries (no parent)

- 孤立摘要（无父节点）

- Dangling raw messages (session mismatch)

- 悬空原始消息（会话不匹配）

- Missing required tables

- 缺失必要表

undefined

undefined

Session Isolation

会话隔离

bash

undefined

bash

undefined

Exclude test sessions from LCM

排除测试会话不被LCM存储

export LCM_IGNORE_SESSION_PATTERNS="test-,temp-,debug-*"

Keep readonly sessions from being stored

设置只读会话不被存储

export LCM_STATELESS_SESSION_PATTERNS="readonly-,audit-"

Restart Hermes

重启Hermes

hermes restart

undefined

hermes restart

undefined

Large Payload Management

大负载管理

bash

undefined

bash

undefined

Enable external storage for large outputs

启用大输出的外部存储

export LCM_LARGE_OUTPUT_EXTERNALIZATION_ENABLED=true export LCM_LARGE_OUTPUT_EXTERNALIZATION_THRESHOLD_CHARS=12000

Enable transcript GC for already-externalized content

启用已外置内容的转录清理

export LCM_LARGE_OUTPUT_TRANSCRIPT_GC_ENABLED=true

Restart Hermes

重启Hermes

hermes restart


External payloads stored in:

~/.hermes/profiles/<profile>/lcm_externalized/

undefined

hermes restart


外部负载存储在：

~/.hermes/profiles/<profile>/lcm_externalized/

undefined

Troubleshooting

故障排除

Plugin Shows as Not Found

插件显示未找到

Symptom:

hermes plugins

shows

lcm (not found)

but tools exist

Solution: If tools are available, LCM is loaded. This is a host discovery mismatch, not a plugin failure.

bash

undefined

症状：

hermes plugins

显示

lcm (not found)

但工具存在

解决方案：如果工具可用，说明LCM已加载。这是主机发现不匹配，并非插件故障。

bash

undefined

Verify tools exist

验证工具存在

hermes chat "Run lcm_status"

If tools work, plugin is functional

如果工具可用，插件功能正常

undefined

undefined

Status Shows Unbound After Restart

重启后状态显示未绑定

Symptom:

/lcm status

shows

session_id: (unbound)

or

threshold_tokens: (uninitialized)

Solution: Send one normal message first:

bash

hermes chat "Hello"
hermes chat "/lcm status"  # Now shows live session data

症状：

/lcm status

显示

session_id: (unbound)

或

threshold_tokens: (uninitialized)

解决方案：先发送一条普通消息：

bash

hermes chat "Hello"
hermes chat "/lcm status"  # 现在会显示实时会话数据

Compaction Not Triggering

压缩未触发

Check threshold calculation:

bash

undefined

检查阈值计算:

bash

undefined

Get current context window

获取当前上下文窗口

hermes chat "What's your context window?"

Calculate expected trigger

计算预期触发值

trigger = context_window * LCM_CONTEXT_THRESHOLD

触发值 = 上下文窗口 * LCM_CONTEXT_THRESHOLD

Example: 128K window, 0.75 threshold = 96K trigger

示例：128K窗口，0.75阈值 = 96K触发值

export LCM_CONTEXT_THRESHOLD=0.75


**Verify compression is enabled**:

```yaml

export LCM_CONTEXT_THRESHOLD=0.75


**验证压缩已启用**:

```yaml

In Hermes config

在Hermes配置中

compression: enabled: true # Must be true

undefined

compression: enabled: true # 必须设为true

undefined

Missing Regex Message Filtering

正则消息过滤缺失

Symptom: Warning about disabled message-level regex filtering

Solution: Install

regex

package:

bash

pip install regex

LCM uses

regex

with timeouts to prevent unbounded pattern matching. Without it,

LCM_IGNORE_MESSAGE_PATTERNS

is disabled.

症状：提示消息级正则过滤已禁用

解决方案：安装

regex

包：

bash

pip install regex

LCM使用带超时的

regex

来防止无限制的模式匹配。如果未安装，

LCM_IGNORE_MESSAGE_PATTERNS

将被禁用。

High Token Costs

令牌成本过高

Tune threshold for your model:

bash

undefined

针对你的模型调优阈值:

bash

undefined

For 1M token model, don't wait until 750K

对于1M令牌模型，不要等到750K才触发

Trigger earlier to reduce costs

提前触发以降低成本

Trigger at 250K (25% of 1M)

在250K时触发（1M的25%）

export LCM_CONTEXT_THRESHOLD=0.25

Trigger at 400K (40% of 1M)

在400K时触发（1M的40%）

export LCM_CONTEXT_THRESHOLD=0.40

undefined

export LCM_CONTEXT_THRESHOLD=0.40

undefined

Database Corruption

数据库损坏

bash

undefined

bash

undefined

Run doctor check

运行doctor检查

export LCM_ENABLE_SLASH_COMMAND=true hermes chat "/lcm doctor check"

Preview cleanup

预览清理

hermes chat "/lcm doctor clean_preview"

Apply cleanup (if safe)

执行清理（如果安全）

export LCM_DOCTOR_CLEAN_APPLY_ENABLED=true hermes chat "/lcm doctor clean_apply"

undefined

export LCM_DOCTOR_CLEAN_APPLY_ENABLED=true hermes chat "/lcm doctor clean_apply"

undefined

Update Plugin

更新插件

bash

undefined

bash

undefined

From plugin directory

从插件目录执行

cd ~/.hermes/plugins/hermes-lcm git pull --ff-only

Or for symlinked install

或符号链接安装的情况

cd /path/to/hermes-lcm ./scripts/update.sh

Restart Hermes

重启Hermes

hermes restart

undefined

hermes restart

undefined

Integration Examples

集成示例

Python Code Using LCM Tools

使用LCM工具的Python代码

python

undefined

python

undefined

Example: Agent helper to search and expand context

示例：Agent辅助工具，用于搜索和展开上下文

async def recover_context(topic: str, max_depth: int = 2): """Search LCM and expand results to recover context."""

# Search for topic
grep_result = await hermes.call_tool("lcm_grep", {
    "pattern": topic,
    "search_raw": True,
    "search_summaries": True,
    "max_results": 5
})

if not grep_result.get("matches"):
    return f"No context found for: {topic}"

# Expand first match
first_match = grep_result["matches"][0]
if "summary_id" in first_match:
    expand_result = await hermes.call_tool("lcm_expand", {
        "summary_id": first_match["summary_id"],
        "max_raw": 10
    })
    return expand_result

return first_match

async def query_history(question: str, scope_summary_id: str = None): """Ask a question about compacted history."""

params = {"query": question, "max_raw": 50}
if scope_summary_id:
    params["summary_id"] = scope_summary_id

result = await hermes.call_tool("lcm_expand_query", params)
return result.get("answer", "No answer generated")

undefined

async def recover_context(topic: str, max_depth: int = 2): """搜索LCM并展开结果以恢复上下文。"""

# 搜索主题
grep_result = await hermes.call_tool("lcm_grep", {
    "pattern": topic,
    "search_raw": True,
    "search_summaries": True,
    "max_results": 5
})

if not grep_result.get("matches"):
    return f"未找到关于以下主题的上下文: {topic}"

# 展开第一个匹配结果
first_match = grep_result["matches"][0]
if "summary_id" in first_match:
    expand_result = await hermes.call_tool("lcm_expand", {
        "summary_id": first_match["summary_id"],
        "max_raw": 10
    })
    return expand_result

return first_match

async def query_history(question: str, scope_summary_id: str = None): """查询压缩历史中的内容。"""

params = {"query": question, "max_raw": 50}
if scope_summary_id:
    params["summary_id"] = scope_summary_id

result = await hermes.call_tool("lcm_expand_query", params)
return result.get("answer", "未生成答案")

undefined

Custom Configuration Template

自定义配置模板

bash

#!/bin/bash

bash

#!/bin/bash

lcm-config.sh - LCM configuration template

lcm-config.sh - LCM配置模板

Core compaction settings

核心压缩设置

export LCM_CONTEXT_THRESHOLD=0.75 export LCM_FRESH_TAIL_COUNT=64 export LCM_LEAF_CHUNK_TOKENS=20000

Dynamic chunking (optional)

动态块大小（可选）

export LCM_DYNAMIC_LEAF_CHUNK_ENABLED=true export LCM_DYNAMIC_LEAF_CHUNK_MAX=40000

Session management

会话管理

export LCM_NEW_SESSION_RETAIN_DEPTH=2 export LCM_IGNORE_SESSION_PATTERNS="test-,temp-" export LCM_STATELESS_SESSION_PATTERNS="readonly-*"

Large payload handling

大负载处理

export LCM_LARGE_OUTPUT_EXTERNALIZATION_ENABLED=true export LCM_LARGE_OUTPUT_EXTERNALIZATION_THRESHOLD_CHARS=12000 export LCM_LARGE_OUTPUT_TRANSCRIPT_GC_ENABLED=false

Model overrides (use env vars for API keys)

模型覆盖（使用环境变量存储API密钥）

export LCM_SUMMARY_MODEL=claude-3-5-sonnet-20241022

export LCM_EXPANSION_MODEL=gpt-4-turbo

Advanced

高级设置

export LCM_ENABLE_SLASH_COMMAND=true export LCM_DOCTOR_CLEAN_APPLY_ENABLED=false

Start Hermes

启动Hermes

hermes start

undefined

hermes start

undefined

Best Practices

最佳实践

Start with defaults: Only tune after observing actual compaction behavior
Monitor token usage: Use
```
lcm_status
```
to track prompt size and compaction triggers
Tune threshold for your model: Don't use 0.75 blindly on 1M token models
Enable slash commands for debugging: Set
```
LCM_ENABLE_SLASH_COMMAND=true
```
during setup
Use source filters: When expanding, filter by relevant files/tools to reduce noise
External payloads for large outputs: Enable externalization if tool results include large JSON/media
Regular doctor checks: Run
```
/lcm doctor check
```
periodically to catch issues early
Session patterns for isolation: Use ignore/stateless patterns to exclude test/debug sessions

从默认值开始：仅在观察到实际压缩行为后再进行调优
监控令牌使用：使用
```
lcm_status
```
跟踪提示大小和压缩触发情况
针对你的模型调优阈值：不要在1M令牌模型上盲目使用0.75的阈值
启用斜杠命令用于调试：设置
```
LCM_ENABLE_SLASH_COMMAND=true
```
以方便设置过程
使用来源过滤：展开时按相关文件/工具过滤以减少干扰
大输出使用外部负载：如果工具结果包含大JSON/媒体内容，启用外置功能
定期运行doctor检查：定期执行
```
/lcm doctor check
```
以提前发现问题
使用会话模式进行隔离：使用忽略/无状态模式排除测试/调试会话

Resources

资源

GitHub Repository
LCM Paper by Ehrlich & Blackman
Hermes Agent
Lossless CLAW (inspiration)

GitHub仓库
LCM论文 by Ehrlich & Blackman
Hermes Agent
Lossless CLAW（灵感来源）

hermes-lcm-context-management

Original

Translation

Hermes LCM Context Management

Hermes LCM 上下文管理

Overview

概述

Installation

安装

Standard Installation

标准安装

General user plugin (all profiles)

通用用户插件（所有配置文件）

Profile-specific install

特定配置文件安装

Symlink Installation

符号链接安装

Profile-specific

特定配置文件

Configuration

配置

Keep compression enabled - LCM needs this gate

保持压缩启用 - LCM需要此开关

Verification

验证

Core Concepts

核心概念

Message Storage

消息存储

Summary DAG

摘要DAG

Bounded Recovery

有限恢复

Environment Configuration

环境配置

Core Settings

核心设置

Compaction trigger (fraction of context window)

压缩触发阈值（上下文窗口占比）

Recent messages protected from compaction

受保护不被压缩的最新消息数量

Raw backlog floor before leaf compaction

叶子压缩前的原始消息最小令牌数

Enable dynamic chunk-sized leaf compaction

启用动态块大小的叶子压缩

Session Management

会话管理

DAG depth retained after /new (-1 all, 0 none)

/new命令后保留的DAG深度（-1表示全部，0表示无）

Exclude sessions from LCM storage (glob patterns)

排除LCM存储的会话（通配符模式）

Keep sessions read-only (glob patterns)

设置会话为只读（通配符模式）

Exclude messages by content regex (comma-separated)

按内容正则排除消息（逗号分隔）

Large Payload Handling

大负载处理

Store oversized payloads externally

将超大负载存储到外部

Compact already-externalized tool results

压缩已外置的工具结果

Model Overrides

模型覆盖

Override summarization model

覆盖摘要模型

Override expansion synthesis model

覆盖扩展合成模型

Timeouts

超时设置

Advanced

高级设置

Custom database path

自定义数据库路径

Enable slash commands

启用斜杠命令

Allow destructive doctor clean operations

允许doctor工具执行破坏性清理操作

Critical pressure bypass (0.0 = disabled)

临界预算压力绕过（0.0表示禁用）

Tuning for Long Context Models