Kore Memory

The memory layer that thinks like a human.

Kore Memory is a Python library that gives AI agents a biologically-inspired memory system. It uses Ebbinghaus forgetting curves to decay memories over time, automatically scores importance without an LLM, and provides semantic search across 50+ languages -- all running fully offline on your machine.

pip install kore-memory

Why Kore Memory?

Most AI memory systems treat storage as a flat key-value dump or require expensive LLM calls for every operation. Kore Memory takes a fundamentally different approach:

• Ebbinghaus Decay -- Memories naturally fade over time, just like human memory. Important memories last longer; trivial ones disappear. Each retrieval strengthens the memory (spaced repetition).
• No LLM Required -- Importance scoring, compression, and search all run locally. No API keys, no cloud calls, no usage fees.
• Fully Offline -- SQLite + local sentence-transformers. Works on air-gapped machines, CI pipelines, and edge devices.
• Multi-Agent Ready -- Namespace isolation via X-Agent-Id headers. Each agent sees only its own memories.

Feature Comparison

Feature	Kore Memory	Mem0	Letta	Memori
Offline operation	Yes	No	No	No
No LLM required	Yes	No	No	Yes
Memory decay (Ebbinghaus)	Yes	No	No	No
Auto-importance scoring (local)	Yes	Via LLM	No	No
Memory compression	Yes	No	No	No
Semantic search (50+ languages)	Yes (local)	Via API	Yes	Yes
Timeline API	Yes	No	No	No
Tags and relations graph	Yes	No	Yes	No
TTL / auto-expiration	Yes	No	No	No
MCP server support	Yes	No	No	No
Batch API (up to 100)	Yes	No	No	No
Export / Import (JSON)	Yes	No	Yes	No
Soft-delete / Archive	Yes	No	No	No
Prometheus metrics	Yes	No	No	No
Web dashboard	Yes	No	No	No

Core Concepts

The Memory Pipeline

Every memory flows through five stages:

1. Save -- Content is received via REST API, Python SDK, or MCP tool
2. Score -- Importance is auto-scored locally on a 1--5 scale (or explicitly set)
3. Embed -- A local sentence-transformer generates a semantic embedding
4. Store -- Memory is persisted to SQLite with an initial decay_score of 1.0
5. Decay -- Over time, the decay score decreases following the Ebbinghaus curve

When you search, results are ranked by an effective score that combines three signals:

effective_score = similarity * decay * importance

This means recent, important, and semantically relevant memories always surface first.

Memory Categories

Memories can be organized into categories:

Category	Use Case
general	Default catch-all
project	Project-specific knowledge
trading	Financial / trading context
finance	Financial records and decisions
person	Information about people
preference	User preferences and settings
task	Tasks, deadlines, schedules
decision	Architectural and strategic decisions

Namespace Isolation

Every API request can include an X-Agent-Id header. Memories are strictly isolated between agents -- agent A cannot read or search agent B's memories. This makes Kore safe for multi-tenant and multi-agent deployments.

Quick Example

# Start the server
pip install kore-memory[semantic]
kore

# Save a memory
curl -X POST http://localhost:8765/save \
  -H "Content-Type: application/json" \
  -H "X-Agent-Id: my-agent" \
  -d '{"content": "User prefers concise responses in Italian", "category": "preference"}'

# Search
curl "http://localhost:8765/search?q=user+preferences&limit=5" \
  -H "X-Agent-Id: my-agent"

Or with the Python SDK:

from kore_memory import KoreClient

with KoreClient("http://localhost:8765", agent_id="my-agent") as kore:
    kore.save("User prefers dark mode", category="preference")
    results = kore.search("dark mode", limit=5)
    for memory in results:
        print(memory.content, memory.decay_score)

Requirements

• Python 3.11+
• SQLite with FTS5 (included in most Python installations)
• No GPU required (CPU-only sentence-transformers)

What's Next?

• Installation -- Install Kore Memory with the extras you need
• Usage Guide -- Save, search, tag, and manage memories
• Configuration -- Environment variables and tuning
• API Reference -- Full REST API documentation
• MCP Server -- Use Kore with Claude, Cursor, and other MCP clients
• Architecture -- How the decay engine, scoring, and compression work