Mechanism-level notes on embedded agent memory, graph structure, query surfaces, and verifiable state changes.
8 min read
Engrava 0.6.0 and the network: what a default install sends, and why auto_embed: false stops embedding on writes while a search still sends non-blank query text to a configured provider unless you supply a query_vector.
engrava · agents · memory · privacy
8 min read
A checklist for agent memory: five things that go wrong once an agent runs long enough, what to ask your own store about each, and an honest account of which ones Engrava fixes, contains, or only makes visible.
engrava · agents · memory · architecture
7 min read
A 3,845-test suite at 94% coverage did not protect six guards, two of its own safety tests, or one field that silently discarded a value users can legitimately set. Mutation testing found all of it.
engrava · testing · mutation-testing · engineering
6 min read
Engrava 0.6.0 scored 81.6% on the full 500-question LongMemEval-S set in August 2026, under the canonical scorer, with no LLM in the memory pipeline. You can re-run it. The July 2026 run on 0.5.0 stays beside it.
engrava · benchmark · memory · agents · retrieval
3 min read
Engrava 0.5.0 moves the MCP server into the standalone engrava-mcp package, now available on PyPI and listed in the MCP Registry as ai.sovantica/engrava.
engrava · mcp · agents · memory
6 min read
Why agent memory needs a first-class record for actions, not just documents, snippets, and conversational facts.
engrava · agents · memory · architecture
6 min read
Deterministic consolidation, a typed graph in SQLite, and an honest look at what agent-memory benchmarks can and can't measure.
engrava · memory · agents · architecture