lore-memory

The git init for AI memory.

Teach your AI once. It remembers forever. Works with every coding agent.

Why

Your AI forgets everything between sessions. You repeat yourself. It makes the same mistakes. lore-memory fixes that.

• One memory, every agent — drop-in MCP server works with Claude Code, Cursor, Windsurf, Codex, and any MCP-compatible tool
• Error recipes — fix a bug once, never fix it again. Pattern-matches stderr automatically with regex + BM25 fallback
• Memory immune system — SHA-256 provenance hashes + trust scoring (user=1.0, agent=0.8, mined=0.6, fleet=0.5) prevent hallucinated facts from poisoning your memory
• Zero cost, fully local — single SQLite file, WAL mode, FTS5 full-text search. No cloud. No API keys. 1 dependency (pyyaml)

Install

pip install lore-memory

Quick Start (60 seconds)

# Store a memory
lore-memory remember "We use PostgreSQL, never MySQL"

# Search memories
lore-memory recall "which database do we use"

# Show statistics
lore-memory stats

# Manage identity context (injected into every session)
lore-memory identity set name=Miles role=CTO project=lore
lore-memory identity get

MCP Server (Claude Code, Cursor, Windsurf)

Add to your Claude Code / Cursor MCP settings:

{
  "mcpServers": {
    "lore-memory": {
      "command": "python3",
      "args": ["-m", "lore_memory.mcp.server"]
    }
  }
}

Or set a custom database path:

{
  "mcpServers": {
    "lore-memory": {
      "command": "python3",
      "args": ["-m", "lore_memory.mcp.server"],
      "env": {
        "LORE_MEMORY_DB": "/path/to/your/memory.db"
      }
    }
  }
}

6 MCP Tools

Example MCP Usage

# In Claude Code (after adding MCP server):
lore_teach("Always use f-strings, never .format() or %")
lore_fix("ModuleNotFoundError: No module named", ["pip install -r requirements.txt", "check venv is activated"])
lore_recall("string formatting convention")
lore_match_procedure("ModuleNotFoundError: No module named 'requests'")

Python API

from lore_memory import LoreMemory

with LoreMemory() as mem:
    # Store memories
    mem.remember("User prefers dark mode")
    mem.remember("Always use async/await", memory_type="fact")

    # Search
    results = mem.recall("theme preference")
    for r in results:
        print(r["content"])

    # Identity context
    mem.identity.set({"name": "Miles", "role": "CTO"})

    # Statistics
    print(mem.stats())

How It Works

Your input
    │
    ▼
┌─────────────┐    SHA-256 hash     ┌──────────────────┐
│  lore_teach │ ──────────────────► │  SQLite (WAL)    │
│  lore_fix   │   trust scoring     │  memories table  │
│  lore_remember                    │  darwin_journal  │
└─────────────┘                     │  darwin_patterns │
                                    │  identity        │
                                    └────────┬─────────┘
                                             │
                                    FTS5 BM25 index
                                             │
    ┌────────────────────────────────────────▼──────────────┐
    │  lore_recall      → ranked results by relevance       │
    │  lore_match_procedure → regex match → FTS5 fallback   │
    └───────────────────────────────────────────────────────┘

Storage: Single ~/.lore-memory/default.db SQLite file. WAL mode for concurrent reads. FTS5 virtual table for BM25 full-text ranking.

Trust scoring: Every memory carries a trust_score (0.0–1.0) and SHA-256 provenance_hash based on content + timestamp. Recall filters by min_trust (default 0.5) to suppress low-confidence memories from polluting results.

Error recipes: lore_fix stores both a darwin_pattern (for fast regex matching) and a memories entry (for FTS5 fallback). lore_match_procedure tries regex first, falls back to BM25 search automatically.

Decay: Every memory has a decay_score (default 1.0). Memories accessed via lore_recall get their access_count incremented. Low-decay memories can be filtered out of search results.

Configuration

Create ~/.lore-memory.yml to override defaults:

db_path: ~/.lore-memory/default.db

layers:
  search:
    top_k: 10
  temporal:
    decay_halflife_days: 30

darwin:
  enabled: true
  pattern_threshold: 3

Config search order: explicit path → ./.lore-memory.yml → ~/.lore-memory.yml → built-in defaults.

vs Others

Ecosystem

License

MIT — see LICENSE