Engrava library uses deterministic memory consolidation to fix long-running AI agents

Developers have released Engrava, an open-source Python library designed to give AI agents structured, inspectable long-term memory stored in a single SQLite file. The library models memory as a typed knowledge graph where nodes represent thoughts and typed edges capture relationships, addressing shortcomings of purely vector-based approaches that can blur or contradict stored information. A deterministic consolidation cycle scores each stored thought across five signals — recency, staleness, confirmation, confidence, and frequency — and only promotes facts that meet configurable thresholds, such as requiring at least two confirmations before a piece of information is treated as settled. The entire consolidation process runs as arithmetic over SQLite rows with no language-model or network calls, making it reproducible and auditable via a YAML policy file. The design draws on cognitive-architecture research, including studies on selective memory stabilization during sleep, to keep the memory system a judgement layer rather than a passive accumulation store.