memory-standard

The dsh ecosystem has dozens of memory plugins but no shared standard: formats are closed, budgets are ad hoc, truncation is silent, and agents cannot read each other notes. memory-standard (mm v1) fills that gap by anchoring memory in plain Markdown files governed by a public SPEC and a JSON Schema, so any compliant reader or writer can operate on the same set of files without a proprietary protocol.

Three pillars drive the design. First, layered memory: MEMORY.md serves as a hand-loaded index with a hard cap (200 lines / 25 KiB by default), while detailed notes live under memories/ and load on demand. Second, deterministic budgets: any write that exceeds its character budget, a note that breaches its byte cap, or an index that passes its hard cap returns a structured error that explicitly demands a rewrite; silent content loss is treated as a data-loss violation. Third, cross-agent interop via mm:/// URIs, a documented Markdown format, and an open read/write API. On top of that, mem_digest ingests official dsh session logs and compaction output to distill candidate memories for human review, entirely LLM-free and never writing automatically.

It suits dsh users who want a portable, auditable memory layer, as well as multi-agent workflows that need deterministic memory contracts. Zero runtime dependencies (Node built-ins only), and it runs both as a dsh plugin and as a standalone CLI.