Orma: Remembering Without Thinking -- Episode-First Cognitive Memory for Local Language Models via Topological Loci and Metric Retrieval
Long-term memory frameworks for LLM agents (e.g., Mem0, Zep, Cognee, LangMem) typically assume cloud-scale resources: frontier models, an LLM call on every memory operation, and graph databases. On commodity laptops running 3B-8B open-weight models, these assumptions cause multi-second stalls, inverted or malformed graph triples, and destructive state updates. We present Orma, an episode-first memory engine for local language models built around a single rule: no LLM calls during retrieval. Orma (i) organises conversations with a computational Method of Loci whose checkpoints are linked by sequential 'walk' trajectories; (ii) modulates candidate scores with ACT-R power-law recency, Hawkes-style burst intensity and Ebbinghaus-style retention; (iii) uses pronoun carry-over heuristics and declarative bracket encoding to compress the injected context; and (iv) retrieves through a deterministic four-stage pipeline of Rocchio feedback, Personalized PageRank, an ONNX TinyBERT cross-encoder and Maximal Marginal Relevance. Extraction of states and events is deferred to an asynchronous background 'sleep' worker. On all 1,536 queries of LoCoMo, zero-LLM Orma reaches 28.17% fuzzy MRR and 46.94% fuzzy Recall@20 with 414 context tokens (a 98.4% reduction relative to full context) at 63.8 ms median latency on a laptop CPU. With sleep consolidation by Qwen-2.5-3B, Answer-in-Context recall rises from 27.21% to 31.71% at 56.2 ms median latency. Orma is open-source under the Apache-2.0 license.
Authors
- Sahil Rane
Institutions
- University of Mumbai (IN)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24421388
- Primary Topic
- Topic Modeling
- Type
- preprint