dsh-memory_rollout

Codex-style per-session memory for DeepSeek Harness (DSH). 中文 README

Early version (0.1.x), 48/48 tests passing, in candidate-release observation. Feedback is welcome.

What problem it solves

Every DSH session starts from zero. When you open a new session, the agent doesn't know what was decided last time, what you prefer, or what pitfalls you already hit. dsh-memory_rollout gives the agent an organized, write-only-when-it-makes-sense persistent memory: facts, preferences, decisions, and project notes survive across sessions, and can be recalled with their source when needed.

Features

• One session, one draft — each qualifying session has its evidence draft generated or appended by Stage 1 (rollout_summaries/<sessionId>.md, like a sub-AGENTS.md); the agent may also explicitly save key points before compaction.
• Layered disclosure — memory_summary.md (injected into the prompt) → MEMORY.md (searchable registry) → a few relevant drafts / notes. Grep-friendly, no full-scan.
• Restrained & passive — automatic generation only processes qualifying sessions; manually adding, editing, or forgetting long-term memory happens only on explicit user request; a quick memory pass (≤4–6 steps) decides when to look memory up, so it never floods the context.
• Idempotent integration — fingerprint + watermark; skips re-integration when nothing changed (no wasted tokens).
• 6 user-facing tools — memory_remember / memory_recall / memory_forget / memory_note / memory_integrate / memory_precompact, plus two memory__* internal scheduler tools.
• Browser management page — a "记忆库 / Memory" page to browse summaries, registry, drafts, and notes; edit config and import/export memory.

Pipeline at a glance

session ends / goes idle → durable queue (Stage 1)
  → extract candidate memories + append-only evidence (draft + source_ref)
  → Phase 2 consolidation + versioned publish (atomic current switch / old versions recoverable)
  → layered read (summary → registry → a few drafts/evidence)
  → remember / forget / supersede enter the unified change stream, re-integrated into an authoritative version

Constraints: no-signal sessions produce no dirty memory; failures never masquerade as success; secrets are redacted at ingress / model / disk; citations point at real content or honestly fall back to unverified; the current user instruction and AGENTS.md take precedence over memory.

Install

dsh plugin --profile web add dsh-memory_rollout

The dsh.bundle manifest wires this plugin into the profile automatically. To install by hand:

pnpm add dsh-memory_rollout

then add a row to your profile's cordis.yml (or cordis.patch.yml):

- id: dsh-memory_rollout
  name: dsh-memory_rollout

Requires a DSH base of 0.1.1-rc.2 or newer (peerDependencies declare ^0.1.1-rc.2). The plugin declares sessionQuery as a required service (provided by the DSH base) — if the base does not mount it, the plugin fails to load and automatic memory (Stage 1 source reading) is disabled.

Usage

Tell the agent to remember something, or do it yourself:

memory_remember(content="用户偏好…", tags=["pref"])   # → long-term memory (with source sessionId)
memory_note(slug="fix-x", content="…")               # → temporary note (only when the user asks)
memory_integrate()                                    # → idempotent integrate summary/MEMORY.md
memory_precompact(content="要留的关键要点")          # → draft + durable queue before compaction

When context from an earlier session matters, the agent runs a quick memory pass: skim the injected summary → search MEMORY.md → open 1–2 relevant drafts → stop if no hits. memory_recall(query="…") is the explicit search entry.

Configuration

The plugin exposes a schemastery config schema. Full parameter table:

Key	Type	Default	Meaning
recallLimit	int	10	Max entries returned by memory_recall
summaryTokens	int	4000	Token budget for the injected memory_summary.md
maxQuickSteps	int	5	Quick-memory-pass search-step budget
memoryRoot	string	''	Override of the memory root; empty = <ds_home>/memories
generateMemories	boolean	true	Whether a session contributes future memory (auto Stage 1); false = no auto enqueue
useMemories	boolean	true	Whether to give the model memory (inject + recall)
maxModelAttemptsPerDay	number	24	Daily Stage 1 model-attempt cap; failed attempts count
extractProvider / extractModel / extractReasoningEffort / maxExtractTokens			Stage 1 extraction LLM route / model / reasoning effort / input-token cap
consolidationProvider / consolidationModel / consolidationReasoningEffort			Phase 2 consolidation LLM route / model / reasoning effort

The settings page can edit these at runtime; changes persist to <ds_home>/dsh-memory_rollout.settings.json and are re-applied on the next startup, taking precedence over cordis.patch.yml. memoryRoot is read-only.

References

This plugin's memory model and LLM extraction prompt are adapted from the memory system of openai/codex (Apache License 2.0) — an independent re-implementation for DeepSeek Harness that does not redistribute its source verbatim. It was also adapted from flymysql/dsh-memory (MIT), the original "cross-session memory vault" it grew out of. See NOTICE for attribution.

License

MIT (see LICENSE).