lore-sdk 0.8.3

Cross-agent semantic memory with knowledge graphs, fact extraction, and MCP integration

Lore — Cross-Agent Memory for AI

Persistent semantic memory that works with every MCP-compatible AI tool.

Agents store what they learn — other agents recall it. Knowledge graphs, fact extraction, auto-consolidation, and more. No API key required for basic use.

Why Lore?

• Local-first — SQLite by default, no server needed. Scale to Postgres + pgvector when ready.
• No API key required — local ONNX embeddings ship with the package. LLM features are opt-in.
• Single database — no Neo4j, Redis, or Qdrant dependency. Everything in one SQLite file or Postgres DB.
• 20 MCP tools — remember, recall, knowledge graph, fact extraction, consolidation, classification, and more.
• Opt-in intelligence — enrichment, classification, fact extraction, and knowledge graphs activate only when you configure an LLM.

Comparison

Feature	Lore	Mem0	Zep	Cognee
Local-first (no server)	Yes	No	No	No
MCP native	Yes	No	No	No
Knowledge graph	Yes	Yes*	Yes	Yes
Fact extraction	Yes	No	No	Yes
Auto-consolidation	Yes	No	Yes	No
Conflict resolution	Yes	No	No	No
Memory tiers	Yes	No	Yes	No
Dialog classification	Yes	No	No	No
Webhook ingestion	Yes	No	No	No
No external DB required	Yes	No**	No	No
PII masking	Yes	No	No	No

* Mem0 requires Neo4j for graph features. ** Mem0 requires Qdrant or Redis.

Comparison as of March 2026. Lore focuses on being the MCP-native, local-first choice for agent memory.

Quick Start

1. Install (30 seconds)

pip install lore-sdk

2. Configure your AI tool (60 seconds)

Add to your MCP client config (e.g., Claude Desktop claude_desktop_config.json):

{
  "mcpServers": {
    "lore": {
      "command": "uvx",
      "args": ["lore-memory"],
      "env": {
        "LORE_PROJECT": "my-project"
      }
    }
  }
}

See Setup Guides for Claude Desktop, Cursor, VS Code, Windsurf, ChatGPT, Cline, and Claude Code.

3. Try it (3 minutes)

Ask your AI assistant:

"Remember that our API uses REST with JSON responses and rate limits at 100 req/min"

Then ask:

"What do you know about our API?"

Lore's recall tool will be invoked automatically.

4. Enable LLM features (optional)

export LORE_ENRICHMENT_ENABLED=true
export LORE_LLM_PROVIDER=anthropic
export LORE_LLM_API_KEY=sk-ant-...

This enables auto-enrichment (topics, entities, sentiment), classification (intent, domain, emotion), and fact extraction on every remember() call.

5. Use the SDK directly

from lore import Lore

lore = Lore()  # zero config — local SQLite, built-in embeddings

lore.remember(
    "Stripe API returns 429 after 100 req/min — use exponential backoff",
    tags=["stripe", "rate-limit"],
    tier="long",
)

results = lore.recall("stripe rate limiting")
for r in results:
    print(f"[{r.score:.2f}] {r.memory.content}")

Full Quick Start Guide

Architecture

graph LR
    A[MCP Client] -->|stdio| B[MCP Server]
    B --> C[Lore SDK]
    C --> D[Store<br/>SQLite / Postgres / HTTP]
    C --> E[Embedder<br/>ONNX local]
    C --> F[LLM Pipeline<br/>optional]
    F --> F1[Enrich]
    F --> F2[Classify]
    F --> F3[Extract Facts]
    F --> F4[Knowledge Graph]
    F --> F5[Consolidate]

Pipeline: remember() → redact PII → embed → store → enrich → classify → extract facts → update graph

Recall: recall() → embed query → vector search → tier weighting → importance scoring → graph boost → return results

Full Architecture Documentation

Features

v0.7.0 — Living Archive

on_this_day · verbatim recall · temporal filters

On This Day surfaces memories from the same calendar day across years. Verbatim Recall returns original words instead of AI summaries. Temporal Filters add date-range filtering to recall (year, month, days_ago, before/after, window presets).

Memory Management

remember · recall · forget · list_memories · stats · upvote · downvote

Core memory operations with semantic search, tier-based TTL (working/short/long), importance scoring, and automatic PII redaction.

Knowledge Graph

graph_query · entity_map · related

Entities and relationships extracted from memories, with hop-by-hop traversal. Graph-enhanced recall surfaces connected memories that pure vector search misses.

Fact Extraction & Conflicts

extract_facts · list_facts · conflicts

Atomic (subject, predicate, object) triples extracted from text. Automatic conflict detection when new facts contradict old ones — supersede, merge, or flag.

Intelligence Pipeline

classify · enrich · consolidate

LLM-powered classification (intent/domain/emotion), metadata enrichment (topics/entities/sentiment), and memory consolidation (merge duplicates, summarize clusters).

Import/Export

ingest · as_prompt · check_freshness · github_sync

Webhook-style ingestion with source tracking. Export memories formatted for LLM context injection. Git-based staleness detection. GitHub issue/PR sync.

Setup Guides

Client	Guide
Claude Desktop	docs/setup-claude-desktop.md
Cursor	docs/setup-cursor.md
VS Code (Copilot)	docs/setup-vscode.md
Windsurf	docs/setup-windsurf.md
ChatGPT	docs/setup-chatgpt.md
Cline	docs/setup-cline.md
Claude Code	docs/setup-claude-code.md

All Setup Guides

Docker

For team setups with Postgres + pgvector:

docker compose up -d

This starts Postgres with pgvector and the Lore HTTP server. Point your MCP client to http://localhost:8765.

# Production (with secure password)
cp .env.example .env  # edit POSTGRES_PASSWORD
docker compose -f docker-compose.prod.yml up -d

Self-Hosted Guide

API Reference

• MCP Tools (20 tools)
• CLI Commands
• SDK Methods
• Environment Variables

Full API Reference

Performance

Operation	Target
remember() no LLM	< 100ms
recall() vector search (100 memories)	< 50ms
recall() vector search (10K memories)	< 200ms
recall() graph-enhanced (2-hop)	< 500ms
Embedding generation (500 words)	< 200ms
as_prompt() 100 memories	< 100ms

Benchmark Results

Migration from v0.5.x

v0.6.0 adds 13 new MCP tools (7 → 20), new database columns and tables, and opt-in LLM features. Existing installations work without changes — all new features are opt-in.

Migration Guide

Examples

See examples/ for runnable scripts:

• full_pipeline.py — remember, recall, tiers, prompt export
• mcp_tool_tour.py — tour of all 20 MCP tool equivalents
• webhook_ingestion.py — ingest with source tracking
• consolidation_demo.py — memory consolidation

Contributing

git clone https://github.com/agentkitai/lore.git
cd lore
pip install -e ".[dev,mcp,enrichment]"
pytest

License

MIT