reeve 0.1.7

Lifetime persistent memory for LLMs. A lightweight SDK and MCP server for structured knowledge graphs.

Reeve

Lifetime persistent memory for LLMs — a knowledge graph that remembers everything, knows what changed, and never forgets what matters.

Reeve is a lightweight SDK and Model Context Protocol (MCP) server for long-term AI memory. It transforms unstructured chat history into a structured Neo4j knowledge graph, ensuring that every fact you store today remains accurately retrievable 50 years from now — in context, and evolving with you.

from reeve import store, query

# Store a memory
store("I just started as a software engineer at Google in San Francisco")

# Months later...
store("I moved from San Francisco to New York")

# Reeve connects the dots across time
print(query("Where do I live?"))        # → "New York."
print(query("Did I ever live in SF?"))  # → "Yes, before moving to New York."

The memory of your move was stored months after your initial job announcement. But when you need it, Reeve connects the dots — because it actually understands the relationship between locations, time, and your career history.

No context windows. No token limits. No forgetting. Persistent memory that outlives any single conversation — or any single year.

---

Why This Exists

LLMs are stateless. Every conversation starts from zero. The common workarounds all break down over time:

Approach	Works for a week	Breaks at scale
Chat history	✓	Grows unbounded, no structure, can't handle contradictions
Vector stores (Pinecone, ChromaDB)	✓	Flat chunks — "I moved to NYC" never invalidates "I live in SF"
RAG pipelines	✓	Retrieves similar text, not relevant knowledge
Context-window managers (MemGPT)	✓	Clever paging, but still raw text — no entity tracking, no state evolution

Ask any of them: "What changed about me since last year?" — silence.

Ask Reeve — it knows, because it actually tracks entities, states, and time.

How It Works

Reeve doesn't store text. It understands it.

Every input is parsed by an LLM into structured semantics — entities, actions, states, roles, emotions, locations — and written into a Neo4j knowledge graph with typed relationships. This isn't an embedding dump. It's a living, evolving model of everything you've told it.

What Happens When You Store a Memory

"I love playing football" becomes a structured graph:

(Episode) ──INVOLVES──▶ (Entity: you)
    │
    ├──HAS_STATE──▶ (State: hobby=football, sentiment=love) ──OF_ENTITY──▶ (Entity: you)
    └──INVOLVES──▶ (Entity: football)

"I just started at Google as a software engineer in San Francisco":

(Episode) ──INVOLVES──▶ (Entity: Google)
    │                        │
    ├──HAS_ACTION──▶ (Action: started working) ──BY_ENTITY──▶ (Entity: you)
    │                                          ──ON_ENTITY──▶ (Entity: Google)
    ├──HAS_STATE──▶ (State: role=software engineer) ──OF_ENTITY──▶ (Entity: you)
    └──AT_LOCATION──▶ (Location: San Francisco)

Months later, when someone asks "My friend invited me to play football, should I go?", Reeve retrieves the football episode by semantic similarity, sees the sentiment=love state, and answers: "Yes! You love playing football."

No explicit link was ever made between the question and the stored memory — the knowledge graph makes the connection automatically.

What Happens When Facts Change

"I moved from San Francisco to New York" — the old state is superseded, not deleted or duplicated. Full history is preserved:

(State: city=New York, active=true) ──SUPERSEDES──▶ (State: city=San Francisco, active=false)

Ask "Where do I live?" → gets the current answer: New York.
Ask "Where was I living when I joined Google?" → still knows it was San Francisco.

This is what persistent memory actually means. Not similarity search — structured, evolving knowledge with a complete audit trail.

Triple-Scoring Retrieval: Why the Right Memory Surfaces

Most systems rank by vector similarity alone. That works for a week. Over years, your "I got married" memory gets buried under thousands of newer, more recent entries.

Reeve scores every candidate across three scoring dimensions simultaneously:

score = 0.65 × vector_similarity + 0.30 × importance + 0.05 × recency

• Vector similarity — semantic relevance via ANN search
• Importance — LLM-assigned at ingestion (landmark life events score higher)
• Recency — exponential decay with a 365-day half-life

The key: memories above importance 0.75 bypass recency decay entirely. Your wedding, your child's birth, a cancer diagnosis — these surface instantly no matter how old they are. Yesterday's lunch order won't outrank them.

Entity Resolution: One Identity, Many Names

"Google", "my company", "the office", "work" — all resolve to the same canonical entity through 3-layer matching:

1. Case-insensitive exact — instant lookup
2. Substring containment — "my company Google" → Google
3. Embedding similarity — cosine ≥ 0.88 catches semantic equivalents

Designed to Last a Lifetime

Most memory systems assume weeks of data. Reeve is engineered for a 70-year lifespan:

What could go wrong	How Reeve handles it
Graph grows to millions of nodes	Dynamic candidate pool scales with size (2% of episodes, 50–500)
Old memories become noise	Consolidation engine LLM-summarizes old low-importance episodes
Entity names fragment over years	Background 3-pass deduplication merges fragmented nodes
Facts become outdated	SUPERSEDES chain — current state always queryable, history preserved
Embedding models get replaced	Model version tags + batch reindex utility
Disaster strikes	Full JSON backup with timestamped exports

Store a memory on day one. Query it on day 25,550 (year 70). It's still there, still accurate, still in context.

---

Features

• Pluggable LLM providers — Ollama (local, free) or OpenAI (cloud); switch via one env var
• Semantic extraction — LLM parses text into structured entities, actions, states, roles, emotions, importance, and location
• Neo4j knowledge graph — Episodes, Entities, Actions, States, Roles, Locations with typed relationships
• 3-layer entity resolution — case-insensitive → substring → embedding similarity (≥ 0.88)
• State contradiction handling — new facts SUPERSEDE old ones with full audit trail
• Vector ANN search — Provider-matched embeddings indexed in Neo4j
• 70-year recency tuning — 5% recency weight, 365-day half-life, importance floor bypass
• Temporal queries — supports "last 3 months", "in 2024", "March 2025" natively
• Memory consolidation — LLM-summarizes old low-importance episodes
• Backup & export — full graph dump to timestamped JSON
• Entity deduplication — 3-pass background scan and merge

Installation

pip install reeve

If you want to run the full memory server locally with MCP support, install MCP extras:

pip install "reeve[mcp]"

Prerequisites

• Python 3.10+
• Neo4j 5.x (Aura or self-hosted with vector index support)
• LLM provider — one of:
  • Ollama (default, free, local) — Install Ollama
  • OpenAI — requires API key from platform.openai.com

Configuration

Copy the example environment file and fill in your credentials:

cp .env.example .env

Option A — Ollama (local, default)

# Install & start Ollama
brew install ollama
ollama serve          # keep running in a separate terminal

# Pull required models
ollama pull llama3.2:3b
ollama pull nomic-embed-text

Your .env only needs Neo4j credentials — Ollama settings default automatically:

LLM_PROVIDER=ollama
NEO4J_URI=neo4j+s://your-instance.databases.neo4j.io
NEO4J_USER=neo4j
NEO4J_PASSWORD=your-password

Option B — OpenAI (cloud)

LLM_PROVIDER=openai
OPENAI_API_KEY=sk-your-api-key
NEO4J_URI=neo4j+s://your-instance.databases.neo4j.io
NEO4J_USER=neo4j
NEO4J_PASSWORD=your-password

Create the required Neo4j vector index (run once in Neo4j Browser — set dimension to match your provider):

-- Ollama (nomic-embed-text): 768 dimensions
-- OpenAI (text-embedding-ada-002): 1536 dimensions
CREATE VECTOR INDEX episode_embedding IF NOT EXISTS
FOR (ep:Episode) ON (ep.embedding)
OPTIONS {indexConfig: {
  `vector.dimensions`: 768,
  `vector.similarity_function`: 'cosine'
}};

Verify your setup:

reeve config

Quick Start

Python API (Local or SDK)

from reeve import store, query

# Store memories
store("I love playing football", speaker="ankesh")
store("I work at Google as a software engineer", speaker="ankesh")

# Query — even months later, across different conversations
answer = query("My friend invited me to play football, should I go?", speaker="ankesh")
print(answer)  # "Yes, you should! You love playing football."

Remote Client (SSE)

If you are communicating with a remote Reeve server:

from reeve import ReeveClient

client = ReeveClient(base_url="https://mcp.reeve.co.in", api_key="your-api-key")
client.store_memory("I love hiking")
print(client.query_memory("What do I love?"))

CLI

# Interactive chat
reeve chat --speaker ankesh

# Store a single memory
reeve store "I love playing football" --speaker ankesh

# Query — connects to stored knowledge automatically
reeve query "My friend invited me to play football, should I go?" --speaker ankesh

# Show active provider, models & run health checks
reeve config

# Admin operations
reeve backup --speaker ankesh
reeve dedup
reeve consolidate --speaker ankesh
reeve reindex --old-model nomic-embed-text

Or via python -m:

python -m reeve chat --speaker ankesh

MCP Server

Run as an MCP server over stdio (recommended for MCP clients like Claude Desktop / VS Code):

reeve-mcp

Network transport (SSE) is also supported:

reeve-mcp --transport sse --host 127.0.0.1 --port 8000

Exposed MCP tools include:

• store_memory
• query_memory
• retrieve_memory_context
• memory_config
• backup_memory
• deduplicate_memory_entities
• consolidate_memory

Deploy MCP Server

Option A — Docker (SSE on port 8000)

Deployment files:

• deploy/mcp/Dockerfile
• deploy/mcp/docker-compose.yml

Run:

cd deploy/mcp
docker compose up -d --build

Stop:

cd deploy/mcp
docker compose down

Client configuration templates for Claude and other MCP clients:

• docs/mcp-client-configs.md
• deploy/mcp/client-configs/

Option B — macOS launchd (always-on local service)

Template file:

• deploy/mcp/com.reeve.mcp.plist

Install and start:

cp deploy/mcp/com.reeve.mcp.plist ~/Library/LaunchAgents/
launchctl load ~/Library/LaunchAgents/com.reeve.mcp.plist

Stop and unload:

launchctl unload ~/Library/LaunchAgents/com.reeve.mcp.plist

Architecture

User Input
  │
  ├── Statement ──▶ operator.py ──▶ reconciler.py ──▶ Neo4j Graph
  │                  (LLM)          (entity resolution,
  │                                   state contradiction,
  │                                   graph writing)
  │
  └── Question  ──▶ retriever.py ──▶ Neo4j Vector Index
                     (triple-scoring)    + Graph Traversal
                    ──▶ llm_interface.py ──▶ Answer
                         (LLM)

See docs/flowcharts.md for detailed architecture diagrams.

Switching Providers

Reeve auto-detects the correct embedding dimension and warns you about mismatches. Run reeve config at any time to check your setup.

Ollama → OpenAI

# 1. Update .env
LLM_PROVIDER=openai
OPENAI_API_KEY=sk-your-api-key

# 2. Install the optional OpenAI dependency
pip install "reeve[openai]"

# 3. Verify (will warn if Neo4j index dimension doesn't match)
reeve config

4. Recreate the vector index (1536-dim for text-embedding-ada-002)

In Neo4j Browser:

DROP INDEX episode_embedding;

CREATE VECTOR INDEX episode_embedding IF NOT EXISTS

FOR (ep:Episode) ON (ep.embedding)

OPTIONS {indexConfig: {vector.dimensions: 1536, vector.similarity_function: 'cosine'}};

5. Re-embed all episodes with the new model

reeve reindex --old-model nomic-embed-text

### OpenAI → Ollama

```bash
# 1. Install & start Ollama
brew install ollama && ollama serve
ollama pull llama3.2:3b && ollama pull nomic-embed-text

# 2. Update .env
LLM_PROVIDER=ollama

# 3. Verify
reeve config

# 4. Recreate the vector index (768-dim for nomic-embed-text)
#    In Neo4j Browser:
#      DROP INDEX episode_embedding;
#      CREATE VECTOR INDEX episode_embedding IF NOT EXISTS
#      FOR (ep:Episode) ON (ep.embedding)
#      OPTIONS {indexConfig: {`vector.dimensions`: 768, `vector.similarity_function`: 'cosine'}};

# 5. Re-embed all episodes
reeve reindex --old-model text-embedding-ada-002

Using a Custom Model

Set these in .env to use any Ollama-compatible model:

OLLAMA_CHAT_MODEL=mistral       # any chat model
OLLAMA_EMBED_MODEL=mxbai-embed-large  # any embedding model
EMBEDDING_DIM=1024              # override auto-detection for unknown models

Advanced Usage

Use the lower-level API for fine-grained control:

from reeve.operator import operator_extract
from reeve.reconciler import reconcile, consolidate
from reeve.retriever import retrieve
from reeve.backup import save_backup
from reeve.entity_dedup import deduplicate_entities

Extract semantics

semantics = operator_extract("Alice presented her paper at MIT")

Write to graph

episode_id = reconcile(semantics, speaker="ankesh", raw_text="...")

Retrieve context

context = retrieve("What did Alice do?", speaker="ankesh")

Admin operations

consolidate("ankesh") save_backup(speaker="ankesh") deduplicate_entities(dry_run=False)

## Project Structure

reeve/ ├── reeve/ # Core package │ ├── init.py # Public API (store, query, close) │ ├── main.py # python -m reeve │ ├── cli.py # CLI entry point │ ├── config.py # Configuration from .env │ ├── schemas.py # TypedDict data contracts │ ├── database.py # Neo4j driver wrapper │ ├── embeddings.py # Embedding wrapper (Ollama / OpenAI) │ ├── llm_interface.py # LLM chat wrapper (Ollama / OpenAI) │ ├── operator.py # Semantic extraction via LLM │ ├── reconciler.py # Graph writer + entity resolution │ ├── retriever.py # Triple-scoring retrieval engine │ ├── entity_dedup.py # Entity deduplication │ ├── backup.py # Graph export to JSON │ ├── chat.py # Interactive chat loop │ └── utils.py # Shared helpers ├── examples/ # Example scripts │ ├── basic_usage.py # Store and query │ ├── advanced_pipeline.py # Lower-level API usage │ ├── streamlit_app.py # Optional Streamlit web UI │ └── debug_retrieval.py # Diagnostic tool ├── tests/ # Test suite ├── docs/ # Documentation │ └── flowcharts.md # Architecture flowcharts ├── pyproject.toml # Package metadata and dependencies ├── .env.example # Environment variable template ├── LICENSE # MIT License ├── CONTRIBUTING.md # Contribution guidelines └── README.md

## Configuration Reference

| Variable | Default | Description |
|---|---|---|
| `LLM_PROVIDER` | `ollama` | LLM backend: `ollama` (local) or `openai` (cloud) |
| `OLLAMA_BASE_URL` | `http://localhost:11434` | Ollama server address |
| `OLLAMA_CHAT_MODEL` | `llama3.2:3b` | Ollama chat model |
| `OLLAMA_EMBED_MODEL` | `nomic-embed-text` | Ollama embedding model (768-dim) |
| `OPENAI_API_KEY` | — | OpenAI API key (required when provider=openai) |
| `OPENAI_CHAT_MODEL` | `gpt-4o` | OpenAI chat model |
| `OPENAI_EMBED_MODEL` | `text-embedding-ada-002` | OpenAI embedding model (1536-dim) |
| `EMBEDDING_DIM` | auto | Override auto-detected embedding dimension |
| `WEIGHT_SIMILARITY` | `0.65` | Vector similarity weight in scoring |
| `WEIGHT_IMPORTANCE` | `0.30` | Importance weight in scoring |
| `WEIGHT_RECENCY` | `0.05` | Recency weight (low for 70-year safety) |
| `RECENCY_HALF_LIFE_DAYS` | `365` | Days for recency to decay by 50% |
| `IMPORTANCE_FLOOR` | `0.75` | Episodes above this ignore recency decay |
| `VECTOR_CANDIDATES_MIN` | `50` | Minimum ANN candidates |
| `VECTOR_CANDIDATES_MAX` | `500` | Maximum ANN candidates |
| `VECTOR_CANDIDATES_RATIO` | `0.02` | Fraction of total episodes to search |
| `CONSOLIDATION_AGE_DAYS` | `90` | Episodes older than this are eligible |
| `CONSOLIDATION_BATCH_SIZE` | `50` | Max episodes per consolidation round |
| `CONSOLIDATION_IMPORTANCE_CAP` | `0.3` | Only consolidate below this importance |

## 70-Year Design Decisions

| Problem | Solution |
|---|---|
| Recency bias buries old memories | Recency weight = 5%, half-life = 365 days |
| Important memories forgotten | Importance floor bypass (≥ 0.75 → recency = 1.0) |
| Fixed candidate pool too small | Dynamic pool: 2% of graph, clamped 50–500 |
| Graph grows unbounded | Consolidation engine merges old trivia |
| State contradictions | SUPERSEDES chain — only latest value active |
| Entity fragmentation | 3-layer entity resolution + background dedup |
| Embedding model changes | Model version tag + batch reindex utility |
| Vector index lag | 5-minute recent-episode safety net |
| No disaster recovery | Full JSON export with incremental backup |

## Development

```bash
# Install with dev dependencies
pip install -e ".[dev]"

# Run tests
pytest tests/

# Lint
ruff check .

# Type check
mypy reeve/

See CONTRIBUTING.md for full guidelines.

Tech Stack

• Python 3.10+
• Neo4j 5.x — Graph database with vector index
• Ollama (default) — Local LLM & embeddings (llama3.2, nomic-embed-text)
• OpenAI (optional) — GPT-4o (chat) + text-embedding-ada-002 (embeddings)
• LangChain — Unified LLM/embedding wrappers
• NumPy — Cosine similarity computation

License

MIT — see LICENSE.