dsh-memories

dsh-memories

中文说明

Dual-ledger cross-session memory for DeepSeek Harness.

Inspired by the memory pipeline of OpenAI's open-source Codex agent, dsh-memories gives every new session in a project two living ledgers:

Both are maintained automatically and recalled into the system context of every new conversation — no re-briefing required.

Features

• Automatic extraction — on each new session (throttled to once per 30 minutes), recent sessions from the same workspace (last 14 days, up to 2 per run) are summarized by an LLM into:
• stable *facts* in four fixed categories: preference / project / environment / lesson
• *progress* movements: completed / doing / next
• Strict no-op gate — one-off task details, code dumps, and secrets are never recorded; sessions with nothing worth keeping produce nothing.
• LLM consolidation — drafts are merged and rewritten into clean, deduplicated, sectioned ledgers (with automatic .bak backups).
• Recall injection — ledger contents ride along in the system prompt of every new session (inline when small, with a file pointer when large).
• Decay — consolidation drops stale entries (30 days unconfirmed) unless marked pinned.
• Model tools & slash commands — remember / update_progress tools let the agent write proactively; /remember, /progress, /memories give humans direct control.
• Failure-safe — LLM failures leave no trace in state; affected sessions are retried automatically next round.

Requirements

A DeepSeek Harness deployment that provides the standard host-plane services:

fs · llm · sessionQuery · systemPrompt · tools · commands · sandboxPolicy

(all ship with the default dsh-base bundle; tested on dsh 0.1.0-rc.9)

Install

Option 0 — one line via dsh plugin add (recommended)

dsh plugin --profile web add dsh-memories

The plugin manager reads this repo's bundled `cordis.patch.yml) and wires everything for you. Needs a release that ships the bundle manifest -- the current npm 0.1.2 predates it, so until the next publish, use Option A.

Option A — npm package into your profile

cd ~/.dsh/profiles/web            # or whichever profile you run
npm install <git-url-or-tarball>  # places dsh-memories into ./node_modules

Then append a row to ~/.dsh/profiles/web/cordis.patch.yml. Reference the entry file by relative path — the proven pattern in pnpm-managed profiles (bare package names in patch rows are not reliably resolved):

- insert:
    - id: dsh-memories
      name: './node_modules/dsh-memories/lib/index.js'

Alternatively, register it npm-natively: add "dsh-memories": "*" to the profile package.json dependencies, add "dsh-memories" to dsh.profile.bundles, run pnpm install, and skip the patch row entirely.

Option B — copy the folder

Copy this repository folder into ~/.dsh/profiles/web/node_modules/dsh-memories, then add the same patch row as above.

Restart DSH once. Done — the plugin loads at boot and works silently afterwards.

If you previously ran this capability as a dynamic plugin (cordis_define/mem-*), remove it before restart to avoid duplicate tool registrations.

Usage

You normally do nothing. For direct control:

Where the data lives

<your-project>/.dsh/memories/
├── MEMORY.md          # consolidated long-term facts (4 sections)
├── MEMORY.md.bak      # previous version, kept on every rewrite
├── PROGRESS.md        # project progress (# 已完成 / ## 进行中 / ## 下一步)
├── PROGRESS.md.bak
└── raw/               # unconsolidated drafts (_manual.md, _progress.md, per-session notes)

Everything is plain Markdown — open it, edit it, diff it.

How it works

new session ──▶ scan (same-workspace sessions, ≤14 days, ≤2/run)
                    │
                    ▼
            extract (one LLM call per session)
            ├─ facts[]      → raw/<session>.md
            └─ progress{}   → raw/_progress.md
                    │
                    ▼
            consolidate (one LLM call per ledger)
            ├─ MEMORY.md    ← merge + dedupe + decay(30d) + sections
            └─ PROGRESS.md  ← 已完成 / 进行中 / 下一步 + dates
                    │
                    ▼
next session ◀── recall section injected into system prompt

The extraction prompt enforces a strict schema (JSON array / NONE), caps entry length, forbids secrets, and skips trivial transcripts (<400 chars). Consolidation prompts enforce section formats and size budgets (≤100 lines for facts, ≤60 for progress).

Limitations & roadmap

• Project-scoped — each workspace keeps its own ledgers. A global user-level ledger (like Codex's ~/.codex/memories) is planned via a globalDir setting.
• Extraction reads only message text (no tool payloads).
• Progress consolidation is instructed — not guaranteed — to respect the 60-line budget.
• Chinese-oriented prompts today; English variants welcome via PR.

License

MIT