caspianmoon/memoripy

What it solves

Memoripy is designed to prevent AI agents from suffering from "memory pollution"—the ingestion of noise, contradictory facts, or untrusted instructions. It solves the problem of maintaining a reliable, evidence-based memory store where the agent can track how facts change over time and explain exactly why a specific piece of information was recalled.

How it works

Memoripy implements a local memory runtime that prioritizes provenance and verification over simple storage. It uses several key mechanisms:

• Admission Barrier: A formal policy that filters incoming data to reject noise, heartbeat signals, or untrusted external instructions before they become durable memories.
• Bitemporal Records: It tracks both when a fact was observed and when it is valid, allowing the system to distinguish between current truth and historical evidence.
• Multi-Lane Retrieval: Instead of relying on a single search method, it combines multiple "lanes" (lexical BM25, semantic similarity, entity overlap, etc.) using reciprocal-rank fusion to ensure the most relevant memory is found.
• Receipts and Citations: Every recalled memory includes a "receipt" explaining the retrieval logic and citations linking back to the original source material.
• Scope Isolation: Memory is strictly partitioned by user, agent, project, or organization to prevent data leakage across boundaries.

Who it’s for

Developers building AI agents that require high-reliability memory, strict data isolation, and the ability to audit the memory state for contradictions or poisoning.

Highlights

• Evidence-First Approach: Prevents feedback loops by refusing to re-ingest retrieved memory as new evidence.
• Local-First: Runs locally with no mandatory third-party dependencies, though it supports external embedding models and databases like Postgres/pgvector.
• Audit CLI: A command-line tool to detect duplicates, conflicting facts, and sensitive data within a memory store.
• MCP Server: Includes an official Model Context Protocol (MCP) server for easy integration with compatible agents.
• Memory Contracts: A vendor-neutral framework to evaluate and benchmark memory behavior against specific test cases.