This OpenCode plugin keeps a coding agent's memory in Markdown you can open and fix

Relevance in this plugin comes down to one frontmatter field. Each block carries a `description` telling the agent how to use it, and without one the agent gets a generic fallback and does not know when the block is relevant ^[6]. The maintainer, who publishes the experimental plugin under the `@ghilteras` scope, says this mirrors the emphasis Letta puts on describing memory blocks well ^[2][6]. So the retrieval rule for every block is prose a human writes and keeps current ^[5]. First run seeds three: `persona` and `human` globally, `project` for the current codebase ^[8].

The journal is off unless you turn it on, and turning it on is one line in `~/.config/opencode/agent-memory.json`: `{"journal": {"enabled": true}}` ^[10]. "I do not want the agent rewriting history; I want it to be able to find what happened last month," the maintainer wrote ^[11]. He is direct about what the log holds. It records what the agent observed, including wrong conclusions, and "Treat it as a log, not a knowledge base" ^[18].

Each `.embedding` sidecar stores `{ v, model, dimension, vector }`, and when the stored dimension does not match the current model the entry degrades to text matching instead of failing the search; bare-array embeddings written by earlier versions still read ^[13]. The in-memory index is per store instance and re-reads an entry only when the `.md` file or its sidecar changes by mtime plus size, so regenerating or deleting a sidecar is picked up on the next search without a restart ^[14]. An edit that leaves both mtime and size unchanged will not be re-read ^[2]. Both details look like the product of something having broken earlier. The model is warmed in the background at plugin init so the first search after a restart does not pay the cold load ^[14].

Since v0.4.2, a query whose text matches an entry title, in either direction and case-insensitively, is guaranteed a high score ^[15]. Titles work as keys, and semantic ranking is what you fall back on when you cannot remember the title ^[6].

The comparison with a flat instruction file is the author's own: he says he relies on AGENTS.md heavily, and describes it as a single document with no notion of scope, no size enforcement and no dedicated operations for the agent to maintain it ^[3]. The failure he is chasing is ordinary. Conventions, debugging decisions and preferences get re-explained after a restart or after context compaction ^[1]. The post reports no evaluation of retrieval quality, and he calls the plugin experimental ^[20]. Each block gets a size limit so it cannot silently grow without bound ^[7], and the caps and descriptions help the agent judge relevance without making the content true ^[17]. "Stale memory is worse than no memory," the maintainer wrote ^[16].