@ruqinhu/dsh-memory

English | 中文

Persistent long-term memory for DeepSeek Harness with configurable character caps.

DSH ships no persistent-memory package, so an agent forgets durable rules between sessions. This plugin keeps two §-delimited documents and injects them as dynamic prompt context on every assembly.

Why not truncate

Truncation is how a memory file quietly goes wrong: the write succeeds, and the next request is missing a constraint. Writes that would exceed the cap are rejected, which forces a real compression pass instead.

The inherited Hermes deployment is the cautionary case — its MEMORY.md had grown to 3,687 characters against a 2,200-character budget (68% over), and its USER.md to 2,748 against 1,375 (100% over).

Install

dsh plugin add @ruqinhu/dsh-memory

The bundle patch is applied automatically via dsh.bundle in package.json.

Configuration

Key	Default	Meaning
memoryCap	2200	Character cap for MEMORY.md
userCap	1375	Character cap for USER.md
enableInjection	true	Inject the documents as dynamic prompt context
backupKeep	12	Backup directories to retain
memoryDir	~/.dsh/memory	Where the documents and backups live
enableRecall	false	Register memory_recall
recallBudgetChars	1500	Character ceiling for one recall

Caps are configuration so a deployment can raise them while inherited files are being compressed, then lower them again.

Injected context

Documents are injected whole while small. Above a 6,000-character combined budget the plugin injects an entry index plus a pointer to memory_search instead, so a growing memory never silently crowds out working context.

Search and recall

CJK is tokenised as bigrams (止损怎么设 → 止损/损怎/么设). This is not a detail. With whole runs kept intact, a query only matched when it happened to be a substring of the entry, so 止损怎么设 scored zero against 止损 2.96/走弱 3.02 and recall returned nothing — indistinguishable from "no relevant memory", which is exactly how the search path had been silently useless. Single characters are still not split out (跌 stays whole), or almost every entry would match.

memory_search returns a scored list of entries for the model to read.

memory_recall (off by default) returns a rendered context block plus one indicator line (🧠 召回 N 条记忆), filled highest-score first until the character budget is spent. Entries that do not fit are skipped, not truncated — half a rule reads like a complete one, which is worse than having none.

Recall is off by default because below the 6,000-character inline budget the whole memory is already in context and a recall would inject a second copy. It pays only once injection has switched to an index. The seam exists now so that change needs no restructuring.

The indicator line follows Hermes' describe_recall, for the same reason: without a deterministic signal, "the model chose to stay silent" and "recall is broken" look identical. This deployment already paid for that lesson on the delivery side.

Tools

Tool	Purpose
memory_read	Read a document with usage and entry count
memory_search	Keyword-search entries; returns scored matches
memory_recall	Search and render a ready-to-use context block (enableRecall)
memory_write	Replace a document (rejected over cap)
memory_append	Append one entry (rejected over cap)
memory_replace	Replace an exact substring in place
memory_maintenance	Back up, rotate, and report usage

memory_replace is the compression primitive: rewrite a stale rule in place rather than rewriting the whole document.

Weekly maintenance

The inherited deployment ran a Saturday compression job. Point a scheduled task at memory_maintenance, then have the agent merge or drop stale entries with memory_replace until usage is back under cap.

License

MIT