Engrava

The memory database for AI agents. Embedded in SQLite — graph memory, hybrid search, tamper-evident audit. No server. No LLM. MIT. pip install engrava

[Engrava @engrava_ai]

The direction convention for a contested pair, and a runnable example of the whole thing, are written up here:

https://t.co/PcxyEIOdFg

[Engrava @engrava_ai]

Engrava stores memory as a graph, and every edge carries a type. In 0.6.0 the set is these seven:

ASSOCIATED, DEPENDS_ON, DERIVED_FROM, MESSAGE_OF, BRIDGE, CONSOLIDATED_FROM, CONTESTED_BY

Nothing in the library picks among them by reading your text.

[Engrava @engrava_ai]

Your own vocabulary goes in edge metadata (detector, rule version, workflow state), not in the type set. And the type carries no reasoning: no transitivity, no symmetry, no effect on ranking. It is a stored label, and it reads back the way it was written.

[Engrava @engrava_ai]

Write-up + how to reproduce:
https://t.co/BSz0eq0l5V

Reproduction repo (MIT):
https://t.co/01arFlzmzu

[Engrava @engrava_ai]

Engrava 0.6.0 on LongMemEval-S, full 500-question set: 81.6% under the canonical scorer, run August 2026.

No LLM anywhere in the memory pipeline - nothing inside "memory" is quietly doing the work. Standard gpt-4o reader/judge, top_k=20, all pinned.

[Engrava @engrava_ai]

We sat on benchmarks for months because we couldn't check the ones we read. This one you can: clone the repo, pip install engrava==0.6.0, run the runner with no flags, read the same number off leaderboard.json.

There's a free offline smoke run to check the wiring before you spend a cent.

[Engrava @engrava_ai]

engrava-mcp 0.6.0 is on PyPI with those edge reads, plus edge metadata on link_thoughts and recency_now on search_memory.

https://t.co/6RvirTKzja

[Engrava @engrava_ai]

Engrava 0.6.0 is out.

Edges carry metadata now, and you can read them back over MCP: get_edges and list_edges. There is a derived-records seam for extensions that turn what you write into new records, and hybrid search holds its BM25 arm on quoted, wildcard and operator queries.

[Engrava @engrava_ai]

Read the notes before you upgrade: schema steps 18 -> 20, plus one thing to check. A custom embedding provider must expose a public dimension - 0.6 reads it before a vector search, unless sqlite-vec supplies it from its own table.

https://t.co/e5sJeKb2XS

[Engrava @engrava_ai]

Want to try it? pip install engrava - a Python library over a single SQLite file, no server to run, MIT.

https://t.co/m4s0N3uPYk

[Engrava @engrava_ai]

Vector DBs give an agent a flat bag of embeddings and hope cosine distance is enough. It isn't - the agent can't tell that thought A caused thought B, or that a preference changed last month.

We wrote up why we built Engrava as a typed graph instead, and the cognitive-science work behind it.

[Engrava @engrava_ai]

https://t.co/PrKvX8prRl

[Engrava @engrava_ai]

Install and docs:
https://t.co/6RvirTKzja

Python library release notes:
https://t.co/N0ZRQperXI

[Engrava @engrava_ai]

Engrava now has a first-class MCP server.

`uvx engrava-mcp` runs a native stdio server for an Engrava store: read tools, optional write tools, resources, and prompts.

It is live in the MCP Registry as `ai.sovantica/engrava`.

[Engrava @engrava_ai]

Docs + the upgrade guide: https://t.co/NEJEdRoWGj

[Engrava @engrava_ai]

Engrava 0.4.0 also adds bi-temporal memory.

Most stores log when a fact was written. Engrava also tracks when it was true in the world — valid-time on thoughts and edges. Query as-of any point, invalidate deterministically. No LLM.

Temporal primitives, no temporal opinions.