Ashish Karigar - MemoryOS

MemoryOS is local-first shared memory for coding agents. When an agent's context compacts or a session ends, the reasons behind the code (decisions, constraints, dead ends) are lost, and the next agent spends time and tokens relearning the project. MemoryOS records those decisions as they happen, links them to the code and the conversation they came from, and hands them back to any agent (Claude Code, Codex, Muse Code) through MCP: a brief at the start of every session and after compaction, and search before changing code. It runs as a Rust daemon with a Mac desktop app; each project's memory stays on your machine. New information extends a decision or flags a conflict instead of silently overwriting it, and every inferred link can be reviewed and undone. Results from our evaluations (same model with and without MemoryOS, 14 tasks x 3 runs): 39/42 trials passed vs 26/42, median 31k vs 58k uncached tokens, 6 vs 10 tool calls. On history-only tasks, where the answer was no longer in the code: 12/12 with memory, 0/12 without. LongMemEval-S (407 held-out questions): 86.7% answer accuracy, evidence in the top 5 for 99% of questions. Search: 45 ms median over 1 million chunks.