local-persona-memory 0.2.0

Long-term personalized memory for ANY local LLM via Ollama + ChromaDB + LangChain

local-persona-memory

Long-term personalized memory for ANY local LLM running via Ollama.

Your AI assistant forgets everything when the session ends. This library fixes that — permanently, privately, and with zero cloud dependency.

What it solves

Most RAG libraries are built for searching documents. This library is built for remembering people. It stores user preferences, facts, skills, and goals, then injects the most relevant memories into every LLM response automatically.

Supported models

Works with any model in the Ollama library:

Use case	Recommended model
General assistant	llama3, llama3.1, mistral
Fast / lightweight	phi3, gemma2:2b, tinyllama
Coding assistant	codellama, deepseek-coder
Multilingual	qwen2, qwen2.5
Reasoning	deepseek-r1, qwen2.5:14b

Install

pip install local-persona-memory

Prerequisites:

1. Install Ollama
2. Pull a model and the embedding model:

   ollama pull llama3
   ollama pull nomic-embed-text
   

3. Run ollama serve (or open the Ollama app)

Quick start

from local_persona_memory import PersonaMemoryManager

# Works with any Ollama model
mem = PersonaMemoryManager(user_id="alice", model="llama3")

# Store memories
mem.remember("Alice is a Python developer building AI tools")
mem.remember("Alice prefers concise answers and code examples")

# Learn from documents
mem.ingest_pdf("notes.pdf")

# Chat with automatic memory context
response = mem.chat("What should I work on today?")
print(response)

# Inspect stored memories
print(f"Stored {mem.memory_count()} memories")

Advanced usage

from local_persona_memory import PersonaMemoryManager, LLMConfig, MemoryConfig, MemoryCategory

mem = PersonaMemoryManager(
    user_id="bob",
    llm_config=LLMConfig(
        model="mistral",         # any Ollama model
        temperature=0.4,
        context_window=8192,     # increase for larger models
    ),
    memory_config=MemoryConfig(
        top_k_memories=8,        # retrieve more memories per query
        chunk_size=512,          # PDF chunk size in characters
        similarity_threshold=0.3,
        embedding_model="nomic-embed-text",
    ),
)

# Categorise memories for better filtering
mem.remember("Bob knows Rust and Go", category=MemoryCategory.SKILL)
mem.remember("Bob wants to build a CLI tool", category=MemoryCategory.GOAL)

# Recall with category filter
skill_memories = mem.recall("programming languages", category=MemoryCategory.SKILL)

How it works

User message
     │
     ▼
[Embed with nomic-embed-text]
     │
     ▼
[ChromaDB similarity search] ──► Top-K relevant memories
     │
     ▼
[Build prompt: system + memories + message]
     │
     ▼
[Generate with Ollama model]
     │
     ▼
Personalised response

Memories persist to ~/.local_persona_memory/ between sessions. Each user gets an isolated ChromaDB collection.

Contributing

Contributions welcome! See CONTRIBUTING.md. Open an issue first for large changes.

License

MIT