Jeshta (ज्येष्ठा)

"I hold your memories so they never fade away~" — Jeshta ✨

Cognitive Memory Engine for Agentic AI

Jeshta is a local-first MCP server that gives LLMs true cognitive memory. Powered by APSM (Active Programmatic Synthesis Memory) — a neuro-symbolic architecture with episodic traces, semantic knowledge graphs, procedural skill learning, and Wake-Sleep consolidation.

---

🧬 APSM Architecture

┌─────────────────────────────────────────────────────────────┐
│                    JESHTA COGNITIVE ENGINE                  │
├─────────────────────────────────────────────────────────────┤
│  ┌─────────────┐   ┌─────────────┐   ┌─────────────┐       │
│  │   LAYER 1   │   │   LAYER 2   │   │   LAYER 3   │       │
│  │  Episodic   │   │  Semantic   │   │ Procedural  │       │
│  │   Stream    │   │   Graph     │   │  Library    │       │
│  │             │   │             │   │             │       │
│  │ Experience  │   │  Knowledge  │   │   Skills    │       │
│  │   Traces    │   │   Triples   │   │    Code     │       │
│  │             │   │             │   │             │       │
│  │ (Context,   │   │ (Subject,   │   │ (Name,      │       │
│  │  Action,    │   │  Relation,  │   │  Code,      │       │
│  │  Outcome)   │   │  Object)    │   │  Execute)   │       │
│  └──────┬──────┘   └──────┬──────┘   └──────┬──────┘       │
│         │                 │                  │              │
│         └────────────┬────┴──────────────────┘              │
│                      │                                      │
│              ┌───────▼───────┐                              │
│              │  WAKE-SLEEP   │                              │
│              │ Consolidation │                              │
│              │               │                              │
│              │ Pattern→Skill │                              │
│              └───────────────┘                              │
└─────────────────────────────────────────────────────────────┘

The Three Layers

Layer	Name	Purpose	Analogy
1	Episodic Stream	Raw experience traces	Hippocampus
2	Semantic Graph	Knowledge relationships	Neocortex (Declarative)
3	Programmatic Library	Executable skills	Neocortex (Procedural)

Wake-Sleep Cycle

• Wake (Active): Log experiences → Query knowledge → Execute skills
• Sleep (Consolidate): Find patterns → Abstract into skills → Prune old data

---

✨ Key Features

Feature	Description
🧠 Tri-Layer Memory	Episodic traces + Semantic graph + Procedural skills
🔄 Wake-Sleep Cycle	Automatic pattern→skill consolidation
🌐 Knowledge Graph	Entity-Relation-Entity triples with Cypher-like queries
⚡ Skill Execution	Store and run Python code safely (AST sandboxed)
📊 Experience Logging	Context, Action, Observation, Outcome traces
🔌 Zero Dependencies	Just Python + SQLite. Local-first, no cloud

---

🛠️ 15 Cognitive Tools

Meta & Session

Tool	Purpose
compile_context	Initialize session, get cognitive state
apsm_status	Memory health check
verify_integrity	Data corruption check

Layer 1: Episodic Memory

Tool	Purpose
log_episode	Log experience (context, action, outcome)
recall_episodes	Search past experiences

Layer 2: Semantic Graph

Tool	Purpose
add_atom	Save fact/preference/constraint
search_atoms	Search knowledge
add_fact	Add relationship triple
query_graph	Cypher-like graph query
recall_related	Get connected nodes
delete_atom	Remove knowledge

Layer 3: Procedural Skills

Tool	Purpose
add_skill	Store executable Python code
execute_skill	Run stored skill
list_skills	List available skills

Wake-Sleep

Tool	Purpose
consolidate	Run pattern→skill cycle

---

🚀 Quick Start

Installation

pip install numpy fastembed

Running

python server.py

MCP Configuration

Add to your MCP config:

{
  "mcpServers": {
    "Jeshta": {
      "command": "python",
      "args": ["D:\\brain\\Jeshta\\server.py"]
    }
  }
}

---

🧪 Usage Examples

Save a Preference (Layer 2)

mcp_Jeshta_add_atom(
    content="User prefers TypeScript over JavaScript",
    intent_mask=4  # Constraint
)

Log an Experience (Layer 1)

mcp_Jeshta_log_episode(
    action="Fixed authentication bug",
    context={"file": "auth.py", "error": "token expired"},
    observation="Token refresh was missing",
    outcome={"success": True, "solution": "Added refresh logic"},
    surprise_score=0.7
)

Add Knowledge Triple (Layer 2)

mcp_Jeshta_add_fact(
    subject="React",
    relation="is_a",
    object="Frontend Framework"
)

Store a Skill (Layer 3)

mcp_Jeshta_add_skill(
    name="format_json",
    description="Pretty print JSON",
    code="import json; result = json.dumps(data, indent=2)",
    parameters=[{"name": "data", "type": "dict"}]
)

---

📁 Files

File	Purpose
apsm.py	Core cognitive engine (Tri-Layer + Wake-Sleep)
server.py	MCP server interface
AI_INSTRUCTIONS.md	How AI should use Jeshta
apsm.db	SQLite database (auto-created)

---

Jeshta — Cognitive memory that learns, forgets, and grows.
Built with ❤️ for the Agentic Future.