Hermes Agent Cognitive Architecture Blueprint

🖥 System State

Runtime

Hermes v0.16.0

2026.6.5 · upstream c0689809

Python

3.11.2

PEP 668 managed

Node.js

v22.23.0

OpenAI SDK

2.24.0

Primary Model

mimo-v2.5

Provider: opencode-go

Base URL

https://opencode.ai/zen/go/v1

Disk Usage

/dev/vda1

33G / 40G 84%

6.6G available

⚙ Configuration — Post-Optimization

Parameter	Previous	Current	Change
memory.memory_char_limit	2,200	4,000	+82%
memory.user_char_limit	1,375	2,000	+45%
compression.threshold	0.91	0.70	−0.21
compression.target_ratio	0.20	0.30	+0.10
tool_output.max_bytes	50,000	100,000	+100%
tool_output.max_lines	2,000	4,000	+100%
context_file_max_chars	null	50,000	new
delegation.model	inherit	kimi-k2.6	pinned
smart_model_routing	—	enabled	new
MCP tools (total)	48	33	−31%

👤 Profiles

Profile	Model	Provider	Status
default	mimo-v2.5	opencode-go	● RUNNING
falcon	kimi-k2.5	opencode-go	○ STOPPED
nova	deepseek-v4-pro	opencode-go	○ STOPPED

🔌 MCP Servers

Server	Transport	Tools	Status
deep-thinker	stdio	8 / 17 filtered	ACTIVE
siyuan_bundinha	stdio	17 / 17 all	ACTIVE
stitch	HTTP	8 / 14 filtered	ACTIVE
elenhub	—	—	REMOVED

deep-thinker — Tool Registry

think evaluate metacog optimize_prompt simulate_devils_advocate conclude reset help

siyuan_bundinha — Tool Registry

unified_search get_document_content create_document append_to_document update_document append_to_daily_note move_documents get_document_tree list_notebooks get_recently_updated_documents create_snapshot list_snapshots rollback_to_snapshot list_all_tags batch_replace_tag list_prompts get_prompt

stitch — Tool Registry

create_project get_project list_screens generate_screen_from_text edit_screens generate_variants create_design_system apply_design_system

elenhub — Removed

DATABASE CORRUPTION

elenhub was removed from the MCP registry after repeated connection failures. The underlying database file was missing or corrupted.

🔑 Credential Pool — opencode-go

Entry	Key Identifier	Status	Role
1	OPENCODE_GO_API_KEY	ACTIVE	Primary
2	api-key-3	ACTIVE	Failover
3	api-key-4	ACTIVE	Failover
—	api-key-2	REMOVED	Stale (401 every session)

Available Models (20)

minimax-m3 minimax-m2.7 minimax-m2.5 kimi-k2.7-code kimi-k2.6 kimi-k2.5 glm-5.2 glm-5.1 glm-5 deepseek-v4-pro deepseek-v4-flash qwen3.7-max qwen3.7-plus qwen3.6-plus qwen3.5-plus mimo-v2-pro mimo-v2-omni mimo-v2.5-pro mimo-v2.5 hy3-preview

🤖 Auxiliary Models

Role	Model	Provider	Usage
Approval	deepseek-v4-flash	opencode-go	User action approval gates
Compression	deepseek-v4-flash	opencode-go	Session history compression
Curator	kimi-k2.6	opencode-go	Content curation & ranking
Kanban Decomposer	kimi-k2.6	opencode-go	Task decomposition
MCP	minimax-m3	opencode-go	MCP tool orchestration
Profile Describer	deepseek-v4-flash	opencode-go	Profile metadata generation
Skills Hub	kimi-k2.6	opencode-go	Skill registry operations

🧠 Honcho Status

ENABLED

Workspace: hermes

Parameter	Value
API Key	`...3j1zq5f4`
Peer Card	34 facts
Recall Mode	hybrid
Dialectic Cadence	every 1 turn
Reasoning	base=low, cap=high
Observation	user+ai, directional
Write Freq	async
Context Budget	uncapped
Hosts	`hermes`, `hermes.falcon`, `hermes_nova` (cleaned from 5)

📐 Architecture Diagrams

1. Memory Hierarchy — 5-Layer Retrieval Stack

L0 → L4 priority order, highest context density at top

┌─────────────────────────────────────────────────────────────────────────┐ │ MEMORY HIERARCHY │ ├─────────────────────────────────────────────────────────────────────────┤ │ │ │ ┌──────────────────────────────────────────────────────────────────┐ │ │ │ L0: SOUL.md (Minimal Kernel) │ │ │ │ ──────── ~2000 chars, byte-stable, identity & core rules │ │ │ │ Always injected, never compressed, never truncated │ │ │ └──────────────────────────────────────────────────────────────────┘ │ │ │ │ │ ▼ │ │ ┌──────────────────────────────────────────────────────────────────┐ │ │ │ L1: USER.md (User Preferences & Profile) │ │ │ │ ──────── memory.user_char_limit: 2000 chars │ │ │ │ Injected per-session, updated via observation │ │ │ └──────────────────────────────────────────────────────────────────┘ │ │ │ │ │ ▼ │ │ ┌──────────────────────────────────────────────────────────────────┐ │ │ │ L2: Honcho (Persistent Peer Memory) │ │ │ │ ──────── Peer card: 34 facts, hybrid recall │ │ │ │ Auto-injected, dialectic every 1 turn, async writes │ │ │ │ Hosts: hermes, hermes.falcon, hermes_nova │ │ │ └──────────────────────────────────────────────────────────────────┘ │ │ │ │ │ ▼ │ │ ┌──────────────────────────────────────────────────────────────────┐ │ │ │ L3: Session Search (Runtime Memory) │ │ │ │ ──────── memory.memory_char_limit: 4000 chars │ │ │ │ Compression threshold: 0.70, target ratio: 0.30 │ │ │ │ Session-local, compressed on threshold │ │ │ └──────────────────────────────────────────────────────────────────┘ │ │ │ │ │ ▼ │ │ ┌──────────────────────────────────────────────────────────────────┐ │ │ │ L4: SiYuan (Persistent Knowledge Base) │ │ │ │ ──────── 17 tools, unified_search, snapshots, tags │ │ │ │ Long-term document storage, notebook organization │ │ │ └──────────────────────────────────────────────────────────────────┘ │ │ │ │ Flow: User message → L0+L1 injected → Honcho recall → L3 context │ │ → L4 on-demand search (unified_search) │ └─────────────────────────────────────────────────────────────────────────┘

2. MCP Tool Flow

Request routing from user intent to MCP tool execution

┌──────────────────────────────────────────────────────────────────────────┐ │ MCP TOOL FLOW │ ├──────────────────────────────────────────────────────────────────────────┤ │ │ │ ┌──────────┐ ┌──────────┐ ┌────────────────┐ │ │ │ USER │────▶│ ROUTER │────▶│ MCP REGISTRY │ │ │ │ INTENT │ │ │ │ (33 tools) │ │ │ └──────────┘ │ smart_ │ └───────┬────────┘ │ │ │ model_ │ │ │ │ │ routing │ │ │ │ └────┬─────┘ │ │ │ │ │ │ │ ┌───────────┼───────────┐ │ │ │ ▼ ▼ ▼ │ │ │ ┌───────────┐ ┌─────────┐ ┌────────┐ │ │ │ │ deep- │ │ siyuan_ │ │ stitch │ │ │ │ │ thinker │ │ bundin. │ │ │ │ │ │ │ 8 tools │ │ 17 tools│ │ 8 tools│ │ │ │ └─────┬─────┘ └────┬────┘ └───┬────┘ │ │ │ │ │ │ │ │ │ ▼ ▼ ▼ ▼ │ │ ┌──────────┐ ┌──────────┐ ┌──────────┐ │ │ │ think │ │ search │ │ create │ │ │ │ evaluate │ │ docs │ │ project │ │ │ │ metacog │ │ snapshot │ │ screens │ │ │ │ ... │ │ tags │ │ design │ │ │ └──────────┘ └──────────┘ └──────────┘ │ │ │ │ Tool Filtering: │ │ deep-thinker: 17 → 8 (47% filtered) │ │ stitch: 14 → 8 (43% filtered) │ │ siyuan: 17 → 17 (0% filtered) │ │ Total: 48 → 33 (31% reduction) │ └──────────────────────────────────────────────────────────────────────────┘

3. Model Selection Decision Tree

Smart routing: cheap model for trivial, default for complex

┌──────────────────────────────────────────────────────────────────────────┐ │ MODEL SELECTION DECISION TREE │ ├──────────────────────────────────────────────────────────────────────────┤ │ │ │ User Request │ │ │ │ │ ▼ │ │ ┌─────────────────────┐ │ │ │ Classify Complexity │ │ │ └─────────┬───────────┘ │ │ │ │ │ ┌──────┴──────┐ │ │ ▼ ▼ │ │ TRIVIAL COMPLEX │ │ (approval, (code gen, │ │ compression, analysis, │ │ routing) reasoning) │ │ │ │ │ │ ▼ ▼ │ │ ┌──────────┐ ┌──────────┐ │ │ │ CHEAP │ │ DEFAULT │ │ │ │ MODEL │ │ MODEL │ │ │ └────┬─────┘ └────┬─────┘ │ │ │ │ │ │ ▼ ▼ │ │ deepseek-v4 mimo-v2.5 │ │ -flash@ocgo @opencode-go │ │ │ │ Auxiliary Model Assignments: │ │ ┌──────────────────────────────────────────────────────┐ │ │ │ approval ──────▶ deepseek-v4-flash │ │ │ │ compression ───▶ deepseek-v4-flash │ │ │ │ curator ───────▶ kimi-k2.6 │ │ │ │ kanban Decomp ─▶ kimi-k2.6 │ │ │ │ mcp ───────────▶ minimax-m3 │ │ │ │ profile desc ──▶ deepseek-v4-flash │ │ │ │ skills_hub ────▶ kimi-k2.6 │ │ │ └──────────────────────────────────────────────────────┘ │ │ │ │ Delegation: kimi-k2.6 (pinned, was inherit) │ └──────────────────────────────────────────────────────────────────────────┘

4. Skills Lifecycle

Full pipeline from observation to promoted skill

┌──────────────────────────────────────────────────────────────────────────┐ │ SKILLS LIFECYCLE │ ├──────────────────────────────────────────────────────────────────────────┤ │ │ │ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ │ │ │ OBSERVE │──▶│ PATTERN │──▶│ DRAFT │──▶│ REPLAY │ │ │ │ │ │ │ │ │ │ │ │ │ │ Watch │ │ Detect │ │ Write │ │ Re-run │ │ │ │ user │ │ repeat- │ │ skill │ │ on new │ │ │ │ behavior │ │ able │ │ skeleton │ │ input to │ │ │ │ │ │ actions │ │ │ │ validate │ │ │ └──────────┘ └──────────┘ └──────────┘ └────┬─────┘ │ │ │ │ │ ┌────────────────────────────┘ │ │ ▼ │ │ ┌──────────┐ ┌──────────┐ ┌──────────┐ │ │ │ VERIFY │──▶│ GATE │──▶│ PROMOTE │ │ │ │ │ │ │ │ │ │ │ │ Assert │ │ LLM │ │ Write to │ │ │ │ output │ │ review: │ │ skill_ │ │ │ │ matches │ │ safety, │ │ manage │ │ │ │ expected │ │ quality, │ │ CRUD │ │ │ │ │ │ utility │ │ registry │ │ │ └──────────┘ └──────────┘ └──────────┘ │ │ │ │ Tools: skill_manage (CRUD) · class-level umbrella · skill hub │ └──────────────────────────────────────────────────────────────────────────┘

📋 Diagnosis Logs

Credential Rotation

2026-06-20 00:21:07,076 credential pool: marking api-key-2 exhausted (status=401), rotating
2026-06-20 00:21:07,076 credential pool: rotated to OPENCODE_GO_API_KEY

MCP Server Registration

2026-06-20 02:11:08,433 MCP server 'deep-thinker' (stdio): registered 8 tool(s)
2026-06-20 02:11:09,484 MCP server 'stitch' (HTTP): registered 8 tool(s)
2026-06-20 02:11:10,776 MCP server 'siyuan_bundinha' (stdio): registered 17 tool(s)
2026-06-20 02:11:10,777 MCP: registered 33 tool(s) from 3 server(s)

elenhub Database Failure (pre-removal)

Error: error returned from database: (code: 14) unable to open database file

Memory Overflow (pre-fix)

Tool memory returned error: Memory at 7,764/2,200 chars. Adding this entry would exceed the limit.

📊 Gap Analysis

Feature	Status	Notes
SOUL.md minimal kernel	✅ Implemented	L0-L2, ~2000 chars, byte-stable
Skills system	✅ Implemented	skill_manage CRUD, class-level umbrella
MCP tool filtering	✅ Applied	tools.include/exclude, 48 → 33 tools
Memory hierarchy	✅ Implemented	5 layers, Honcho auto-injection
Credential pools	✅ Working	3 entries, auto-rotation
Smart model routing	✅ Enabled	deepseek-v4-flash for trivial tasks
Compression tuning	✅ Applied	threshold 0.70, target 0.30
soul.d/ loader	❌ Not recommended	Would break prompt caching
Intent → plugin routing	❌ Requires upstream	Needs NLP classification
Skills CI/CD gates	❌ Requires upstream	Needs test harness
Complexity-aware routing	⚠️ Partial	Beyond cheap/default split

🔬 Deep Thinker State

Binary

/root/.local/bin/deep-thinker-hermes-bundinha

Profile

hermes-bundinha

Saved Sessions

All with 0 nodes — were reset

Autosave

14 KB

AIGuaratuba civic dashboard audit

Current Session Usage

9 nodes 0.87 confidence

Used for gap analysis — 9 reasoning nodes generated with high confidence score

📦 Available Models — Full Registry

#	Model	Provider	Use
1	minimax-m3	opencode-go	MCP orchestration
2	minimax-m2.7	opencode-go	General
3	minimax-m2.5	opencode-go	General
4	kimi-k2.7-code	opencode-go	Code tasks
5	kimi-k2.6	opencode-go	Delegation, curator, skills
6	kimi-k2.5	opencode-go	falcon profile
7	glm-5.2	opencode-go	General
8	glm-5.1	opencode-go	General
9	glm-5	opencode-go	General
10	deepseek-v4-pro	opencode-go	nova profile
11	deepseek-v4-flash	opencode-go	Cheap model (routing)
12	qwen3.7-max	opencode-go	General
13	qwen3.7-plus	opencode-go	General
14	qwen3.6-plus	opencode-go	General
15	qwen3.5-plus	opencode-go	General
16	mimo-v2-pro	opencode-go	General
17	mimo-v2-omni	opencode-go	General
18	mimo-v2.5-pro	opencode-go	General
19	mimo-v2.5	opencode-go	DEFAULT
20	hy3-preview	opencode-go	General

open-source · humanitarian · community-driven

cada build é um passo·cada doação é um impulso