MnemosyneOS/mnemosyne

mnemosyne OS is a zero-third-party-dependency, local-first AI memory engine that installs via pip install mnemosyne-os and runs as a Python library, a CLI, an HTTP API, or a Model Context Protocol server, exposing 31 MCP tools together with a built-in web console and REST endpoints on /v1, /v2, /v3. The core ships with graph memory, multimodal ingestion, reranking, temporal reasoning, a hash-chained audit ledger, and AIC lossless compression, while 72 optional provider adapters cover 20 LLM vendors, 13 embedders, 28 vector stores, six graph stores, and five rerankers — including OpenAI, Qdrant, and Cohere — letting teams swap cloud or self-hosted backends without rewriting application code. Integration guidance with DeepSeek Harness and other agent clients lives in docs/DEPLOY_DEEPSEEK_HARNESS.md, docs/COMPATIBILITY.md, and CHANGELOG.md.

It is aimed at individual developers, agent engineers, and teams that need persistent, auditable memory under local-first or air-gapped constraints, because the built-in rule extractor and offline embedder let Memory() boot, write, and search without any API key, model download, or database installation. A typical workflow begins with mnemosyne init to create a brain directory, uses mnemosyne add to store facts and mnemosyne search --user-id for filtered recall, or drives the engine programmatically through Memory, MemoryClient, or the MemoryBrain facade that returns cost-aware envelopes such as results, cost = brain.recall("...", k=5, budget_tokens=100). When an LLM is added via Memory.from_config, missing credentials automatically fall back to the rule-based engine and surface the reason inside m.describe()["degraded"], so nothing degrades silently.

Runtime requirements are Python 3.8 or newer on a laptop, server, or serverless host, with the core package declaring an empty install_requires so nothing is pulled in unless the user opts into a provider. The repository itself is tagged NOASSERTION, but the README and PyPI badge state that the license is MIT, and users should inspect LICENSE directly before redistribution. First-run users are advised to run python verify.py for the self-check, then scripts/verify_api.py, scripts/verify_memory_lifecycle.py, scripts/verify_precision_recall.py, and scripts/verify_recall_quality.py for offline API, lifecycle, precision, and recall regression coverage, and to read docs/ACCEPTANCE_GUIDE.md, docs/RECALL_STRATEGY.md, and docs/KNOWN_DEFECTS.md to understand the recall strategy and confirmed defects before production deployment.