dsh-memory-ga

Gated, file-based long-term memory for DeepSeek Harness — inspired by GenericAgent’s discipline: No Execution, No Memory.

Your agent already has Skills.
What it usually lacks is a small, trustworthy index of facts + rules, a session notepad, and a settlement ritual that turns hard-won lessons into Skills / facts — without silently poisoning memory from every chat.

Language: English · 中文

---

Why this exists

Coding agents forget. Vector “auto-memory” products remember too much — drafts, wrong turns, guesses.

dsh-memory-ga takes the other path:

Principle	What you get
Verified only	Long-term writes only after tools (or the user) confirmed reality
Hard-injected L1	A ≤~30-line index + RULES every turn — not a pointer you might never open
Working notepad	Session key_info re-injected while non-empty
Gated settlement	start_long_term_update opens a protocol; it does not auto-edit your files
Skills stay Skills	Reusable procedures → DSH Skills. Memory is not a second SOP warehouse
Local & auditable	Plain UTF-8 under $DSH_HOME/memory — diff, backup, hand-edit

No embeddings required. No cloud bank. No silent retain.

---

Architecture (deliberately small)

$DSH_HOME/memory/
  L0_memory_management.md   # constitution (how to remember)
  global_mem_insight.txt    # L1: NAV + RULES  → hard-injected
  global_mem.txt            # L2: verified facts → read on demand
  .working/<session>.txt    # optional session notepad dump

DSH platform (not this folder)
  Skills / skill catalog    # procedures
  Sessions                  # full transcripts (query via DSH, not a memory “L4” tree)

There is no L3 notes library and no memory-tree L4.
If a Skill is wrong, fix the Skill — don’t keep a shadow errata page in memory.

Verified + worth keeping?
  ├─ reusable workflow  → DSH Skill (+ optional L1 one-liner)
  ├─ env / config fact  → L2 (± L1 pointer)
  ├─ one global rule    → L1 RULES
  └─ this task only     → Working (or discard)

---

Tools

Tool	Role
update_working_checkpoint	Replace session notepad (key_info)
start_long_term_update	Return L0 + settlement protocol (no automatic writes)
memory_status	Paths, L1 size, working empty?, nudge counters

Optional soft nudges (default on): after enough steps, gently remind checkpoint / settlement. Never force tool calls. Never auto-write L1/L2.

---

Install (DeepSeek Harness profile)

Compatible with DSH Cordis plugins (dsh.bundle.patch).

A. From Git (recommended once published)

In your profile directory (example: $DSH_HOME/profiles/web):

pnpm add dsh-memory-ga@github:DiligenceLai/dsh-memory-ga

Ensure the profile loads the package (either list it under dsh.profile.bundles or rely on the package’s cordis.patch.yml insert — same pattern as other DSH community plugins).

B. Local path (development)

pnpm add dsh-memory-ga@file:../path/to/dsh-memory-ga

Note: some package managers copy file: deps. After editing the plugin source, reinstall/relink the profile dependency and restart DSH (or wait for profile HMR).

C. Skill helper (optional)

Copy the bundled skill into the user skills root so the agent can load memory-management:

skills/memory-management/SKILL.md  →  $DSH_HOME/skills/memory-management/SKILL.md

Restart or wait until the skill filesystem rescans.

---

Quick verify

1. Restart DSH / reload the profile.
2. In a session, run memory_status → should show $DSH_HOME/memory.
3. update_working_checkpoint with a short note → next model step should include ### [WORKING MEMORY].
4. start_long_term_update → protocol + full L0; disk unchanged until you edit.

First boot creates missing L0/L1/L2 from templates and never overwrites existing files.

---

Configuration

Cordis config on plugin id memory-ga (optional):

- id: memory-ga
  config:
    # root: null            # default: $DSH_HOME/memory
    bootstrap: true
    injectL1: true
    injectWorking: true
    l1MaxChars: 1200
    workingMaxChars: 1200
    persistWorkingFile: true
    nudge:
      enabled: true
      workingEvery: 12
      settleAfterSteps: 15
      maxWorkingNudges: 3
      maxSettleNudges: 2

Hard dependency: the plugin declares inject: ["tools", "systemPrompt", "llm"]. @deepseek-ai/dsh-llm (with createUserMessage) must be present in the host composition — working-memory and nudge injection rely on it. If it is missing, the plugin still loads but memory_status will warn that working/nudges are not injected, rather than silently failing.

workingMaxChars is the single cap for both storing and injecting the working notepad, so memory_status.workingChars always reflects what the model actually sees.

---

What this is not

• Not a vector database / TEMPR / auto-git retain product
• Not a replacement for DSH Skills
• Not automatic Skill generation
• Not a dump of full chat transcripts into memory/

If you want maximum automatic recall across tools, look at other stacks.
If you want controlled crystallization and git-friendly truth, you’re in the right place.

---

Privacy

This repository ships generic templates only.
Your real L1/L2 live under $DSH_HOME/memory on your machine and are never part of this package.
Do not commit personal memory files, tokens, or absolute home paths into forks.

---

Development

git clone <this-repo>
cd dsh-memory-ga
pnpm install   # schemastery + peers for local smoke

Entry: lib/index.js (Cordis name / inject / Config / apply).

---

Contributing

Issues and PRs welcome: install docs for more profiles, stronger nudge hooks, project-scoped memory overlays, session-search Skill examples.

---

License

MIT — see LICENSE.

GenericAgent is a separate project; this plugin only borrows the memory discipline, not GA’s full runtime.