Mnemoria

Cognitive memory system for AI agents — ACT-R activation, typed facts, Hebbian links, RL reranking.

Installation

pip install mnemoria

Quick Start

from mnemoria.store import MnemoriaStore
from mnemoria.config import MnemoriaConfig

store = MnemoriaStore(MnemoriaConfig.balanced())
store.store("V[api.url]: https://example.com")
results = store.recall("What is the API URL?")

For better semantic recall on a real machine, install the embeddings extra:

pip install 'mnemoria[embeddings]'

Note: in constrained containers, TF-IDF fallback is expected and benchmark results may be lower than on a local machine with real embeddings.

Features

• ACT-R Activation — Frequency + recency based activation scoring
• Typed Facts — MEMORY_SPEC notation (C/D/V/?/✓/~) with metabolic decay rates
• Hebbian Links — NPMI-normalized co-occurrence edges with Ebbinghaus decay
• RL Reranking — Q-value UCB-Tuned exploration bonus
• Self-Optimizing Pipeline — LinUCB bandits per retrieval stage
• Scope Lifecycle — active → cold → closed with gauge pressure management
• Contradiction Detection — Entity overlap + update language pattern matching
• IPS Debiasing — Inverse propensity scoring to counteract popularity bias
• PPR Exploration — Personalized PageRank multi-hop discovery

Migration

If you are moving from Honcho or another external memory provider, see:

• MIGRATING_FROM_HONCHO.md

Acknowledgements

If you want the provenance of ideas and inspirations, see:

• ACKNOWLEDGEMENTS.md

License

AGPL-3.0-or-later — see LICENSE file or https://www.gnu.org/licenses/agpl-3.0.html