RESEARCH
What we're testing — and where it has led.
Seven lines of applied research, from memory for AI agents to faithful Hinglish translation. Each one says what we built, what the evidence shows, and what it changed.
We publish a result only when we have a log for it. Work that is not proven yet is labelled Exploring. Dead ends are written up too, because they shaped what we built next.
HOW TO READ THE STATUS
- IN USE1 entryUsed in our own daily work; log on file.
- MEASURED1 entryBenchmarked on real hardware; log on file.
- PROTOTYPE3 entriesWorks; not benchmarked yet.
- EXPLORING2 entriesResearch and design notes.
ALL RESEARCH · 7
- R1IN USEAug 2025–May 2026Long-term memory for AI agentsCan a small, fully local server give AI coding agents a memory that lasts between sessions?AGENT MEMORY · MCP505 · Bootstrap starts
- R2EXPLORINGFeb–Mar 2026Keeping autonomous agents from breaking themselvesWhat stops an autonomous agent from failing in a chain, or from damaging itself when it is allowed to change its own setup?AGENT RELIABILITY45 hours · Outage caused by the failure chain
- R3EXPLORINGEarly 2026State space models for long data streamsCan a Mamba-style state space model act as a cheap, long-memory context model for a fast live data stream?STATE SPACE MODELS · MAMBAabout ±10¹⁵ · Hidden state after 1,024 steps
- R4MEASUREDEarly 2026Real-time ML pipelines on modest hardwareCan a Python-to-C++ pipeline on a laptop and one rented GPU run ML models on a live data stream at 100 Hz without falling over?REAL-TIME ML · SHARED MEMORYsustained 100 Hz with zero crashes · 24-hour paper run, pipeline
- R5PROTOTYPEMar 2026Context that survives between AI coding sessionsCan a plain folder structure, set up by one command, help AI coding assistants keep their context from one session to the next?DEVELOPER TOOLING · NPM4.0.0, on 5 Mar 2026 · Published version
- R6PROTOTYPEMar–Apr 2026Making businesses readable to AI answer enginesCan a store's product catalogue be turned into a dense, factual document that AI crawlers and answer engines can read easily?ANSWER-ENGINE OPTIMISATIONup to 50 · Products read per store
- R7PROTOTYPEMay 2026Faithful Hinglish translation with LLMsCan an LLM translate a long English text into readable Hinglish, sentence by sentence, without adding its own commentary?HINDI & HINGLISH · TRANSLATION136 pages · Source length
METHOD
How we work on research
- 01Measure before claiming.A number goes on this site only when a log backs it. Our agent-memory latency stays unpublished until we re-run the benchmark properly.
- 02Keep a slow LLM out of the hot path.A local 8B model wrote about 3.58 tokens a second on CPU. So our memory server returns passages, and the calling agent writes the answer.
- 03Guardrails before autonomy.After a 45-hour agent blackout, we wrote rules first: agents never edit their own config, and they report a deprecated model instead of fixing it themselves.
- 04Write up the failures.A model trained on 1.4 seconds of data. A hidden state that grew to about ±10¹⁵. A “digital organism” that was really a RAG pipeline. They are all here, because they changed what we built next.