Building YantrikDB: persistent cognitive memory for AI agents. Rust | Python | MCP | server. Apache-2.0. Publishing the benchmarks, failures, and fixes.
Your coding agent did not forget because it was careless. Its decision fell out of the context window.
A fresh agent asks first. YantrikDB returns the decision, source, and why it surfaced.
Apache-2.0: Rust, Python, MCP, server.
https://t.co/NvuILccQlE
mcpier shipped 7 releases in one 2026-04-17 stretch. The last one hung silently -- no error, just SSE frames that never arrived.
49 minutes later: a regex assumed LF line endings against a CRLF upstream. Still no test for that path.
https://t.co/hEJEkkouLs
@AnthropicAI My usage is burning exponentially all of a sudden. The same work pattern from my end as last week. I could literally see usage climbing every couple of mins. I am using Opus 5.5 High.
yantrikdb#256: in September I filed a namespace-check gap as "low priority -- no user affected today." Six weeks later a contributor found its twin: a function with the same gap, no safety net, that had been writing one namespace's text into another's log the whole time.
https://t.co/KB8IMRCRnl
yantrikdb#253: I replied to a bug report with a theory. A day later I partially retracted it -- the real cause was an English-only embedder scoring unrelated Russian text 0.83-0.93 similar. A contributor's fix still leaves one gap open:
https://t.co/X7y9g9nh4e
yantrikdb #251: a contributor found a temp-dir leak, ~490MB over 8 days. I had already written its exact cause into my own PR two months earlier and done nothing about it. Full 18.5-hour writeup, including where my own fix direction went wrong:
https://t.co/UoAxOaE94W
@AlexFreitasAI @victor_explore date alone isnt enough for the conflict. id keep both with a validity window and let last-confirmed plus source weight pick the live one, so the loser stays history instead of getting overwritten
@harisanthosh9 yeah compression alone doesnt help when the repo moved. id invalidate off the files that changed, and keep a short log of what got dropped so you can see why the agent stopped trusting something
I am now using claude as the orchestrator and heavy lifter while Strata is helping to pickup easy ones. So far last 6 days was amazing. Less headache on worrying about usage.
@enhansai@DhravyaShah@supermemory yeah and it's usually not the fact that goes stale, it's the reason behind it. if the memory kept what the decision depended on you could flag it when that changes, instead of it quietly winning forever
@witcheer the part i'd test is whether it comes back. if the fact still lives in old chats the next memory pass can just learn it again. forget probably needs a tombstone, not only a delete