Summon

Self-evolving knowledge graph for AI agents — persistent memory that gets smarter with every interaction.

pip install, 2 lines of config, your agent remembers everything.

Why Summon?

AI agents forget everything between sessions. Vector DBs remember but don't understand relationships. Summon gives agents structured, evolving memory that:

• Self-evolves — frequently used knowledge strengthens; stale knowledge decays

• Connects the dots — automatic relationship detection between memories (graph edges)

• Detects contradictions — flags conflicting memories before they poison your agent's output

• Learns from usage — recall feedback loop tunes retention automatically

Related MCP server: MahoRAGa

Quick Start

Install

pip install summon-mcp

Use as a Python SDK

from summon import Summon

sb = Summon()

# Remember
sb.remember("The production database is PostgreSQL 15 on AWS RDS", tags=["db", "prod"])
sb.remember("API rate limit is 1000 req/min per user", tags=["api", "limits"])

# Recall
results = sb.recall("what database do we use?")
for r in results:
    print(f"[{r.confidence:.0%}] {r.content}")

# Link memories
sb.link(source_id=1, target_id=2, relationship="depends_on")

# Traverse the knowledge graph
graph = sb.traverse(memory_id=1, hops=2)

Use with Claude Code

Add to ~/.claude/claude_desktop_config.json:

{
  "mcpServers": {
    "summon": {
      "command": "python",
      "args": ["-m", "summon"],
      "env": {
        "SUPERBRAIN_DB_PATH": "~/.summon/memory.db"
      }
    }
  }
}

That's it. Claude Code now has persistent memory.

Features

30 MCP Tools

Self-Evolution Engine

• Edge heating — frequently traversed paths strengthen; cold ones decay

• SM-2 spaced repetition — memories reviewed on optimal schedules (like Anki)

• Auto-tune decay — retention thresholds adjust based on actual usage patterns

• Recall feedback loop — tracks which searches were useful, learns from it

• Episodic consolidation — old episodes auto-summarize into permanent facts

Storage

• SQLite — zero-config local storage, perfect for single-user

• Pluggable backends — swap in PostgreSQL, ChromaDB, or custom stores

• BYOM embeddings — bring your own model (OpenAI, DeepSeek, Ollama, local sentence-transformers)

SDK Reference

from summon import Summon, Memory, Edge

sb = Summon()                          # local SQLite (default)
sb = Summon(base_url="...", api_key="...")  # remote API

# Write
mem = sb.remember("fact", tags=["tag"], confidence=0.8)

# Read
mem = sb.get(memory_id)
results = sb.recall("query", mode="hybrid", limit=10)

# Graph
edge_id = sb.link(source=1, target=2, relationship="depends_on")
graph = sb.traverse(memory_id=1, hops=2)

# Manage
sb.reinforce(memory_id)
sb.forget(memory_id, mode="decay")   # or mode="delete"
sb.stats()                           # database statistics
sb.strongest()                       # top memories
sb.weakest()                         # at-risk memories

Community

• License: Apache 2.0 — free for commercial use

• Python: 3.9+

• Status: v0.4.0 Beta — stable for personal use, API may evolve before 1.0

Roadmap

• Cloud sync ($5/mo)

• Multi-tenant SaaS

• LangChain / CrewAI integrations

• Web dashboard

• v1.0 stable API

Built for developers who want their AI agents to stop forgetting.