Blog
Engineering notes on navigation, ingestion, agent memory and honest benchmarking, written by the people who ran the experiments.
We swapped the 12B navigator for a 2.5B one, MiniCPM5-2B in about 3.3 GB of VRAM, and re-ran the benchmark on the same forest: 18/18, 11/11, 7/8. The score did not move. What moved, what did not, and the one miss, explained.
5 min read
The series closer: a self-hosted forest, scoped keys, one row per person, and a locked gate. Twenty days of measured claims and published failures, recapped in three sentences.
3 min read
Ten typed tools over MCP, a skills file that teaches the agent the terrain, and one scoped key per agent: how a self-hosted forest becomes memory your agents can search, walk and grow.
4 min read
No cluster, no A100: the entire benchmark ran on one RTX 3060 with 12 GB. What fits on 12 GB, what to do with 8, and how to run with no local GPU at all through any OpenAI-compatible endpoint.
3 min read
Signals that lexical search over curated metadata will carry you, signals you genuinely want vectors, and a fifty-query gold-set test to settle it on your own corpus before paying for infra.
4 min read
Why agents get ten small, typed, token-budgeted tools over MCP instead of one giant search box, and what that buys in tokens, latency, and legible failures.
4 min read
The engine is Apache-2.0, the apps are AGPL. What each side lets you do, why the split exists, and why the boundary never reaches into your code.
3 min read
Ingest is where the forest spends its effort: 1.71 s per document, a sixty-token summary contract passed at 100%, zero broken links. Pay once at write time, in a batch you can watch, instead of on every query forever.
4 min read
17.5 s p95 for iterative RAG, 8.4 s for the navigator, same 12B model. A dissection of both loops, and why variance measured in cheap hops instead of full retrieve-and-read rounds collapses the tail.
3 min read
The paper carries the full architecture, the benchmark tables and the findings that failed their criteria.