dsh-daoing-memory
Self-evolving memory for DeepSeek Harness (DSH) — earned experiences, diary/fact semantic memory, concern tracking, and an append-only audit ledger.
中文文档 →
An agent without memory starts from zero every session. dsh-daoing-memory gives a DSH agent a persistent, self-improving memory that it earns through use: it keeps a diary, distills durable facts and open concerns about the user, accumulates verified experiences, recalls what is relevant, revises what turned out wrong, and records every change in an auditable ledger.
The design follows four verbs — 生 · 用 · 修 · 记 (Generate · Use · Revise · Record).
Why
Most "memory" bolt-ons either dump raw conversation into a vector store, or let the model write anything it likes into a key-value blob. Both fail in practice: the first buries signal in noise, the second lets a single hallucination poison every future session.
dsh-daoing-memory takes a different stance:
- Memory must be earned. An experience starts as a low-trust candidate and is promoted only after it is corroborated by real use. Nothing reaches high trust by fiat.
- Two distinct memories. Semantic memory (durable facts about the user + open concerns they care about) is separated from experiential memory (how-to knowledge with a lifecycle). They are written, recalled, and governed differently.
- Every write is auditable. An append-only ledger records each mutation, so memory drift or poisoning can be detected, attributed, and rolled back.
- The human stays in the loop. A browser workbench lets you read, correct, promote, and delete memories — memory is a shared artifact, not a black box.
Features
| Area | What you get |
|---|
| Diary (记) | memory_fact — append raw session notes; the substrate everything else is distilled from. |
| Extraction (生) | memory_extract — distill diary entries into durable facts (9 user-centric categories, deduplicated & corroborated) and concerns (todo / thinking / idea / question / decision / commitment, each with a background scene). |
| Experience lifecycle (生·用·修) | memory_ingest, memory_report, memory_revise, memory_refine, memory_verify — experiences are born as candidates, earn trust through reported use, get revised when wrong, and roll back cleanly. |
| Recall (用) | memory_recall — keyword/situation relevance over the shared experience library, with optional context scoping. |
| Audit | memory_ledger, memory_verify — query the append-only ledger and verify integrity. |
| Consolidation | memory_consolidate — periodic compaction/housekeeping of the store. |
| Profile injection | A compact profile snapshot (top facts + open concerns) is injected into the system prompt, so the agent knows the user without being asked. |
| Workbench UI | A browser panel (Fact Diary / Experiences / Ledger / Human Ops) to inspect and curate memory by hand. |
| Extraction skill | A bundled memory-extraction skill that teaches the agent when and how to extract high-quality memory. |
Install
Two DSH environments are supported — a source checkout of DSH and an officially installed DSH — and two install channels (npm package name, or a git/GitHub URL). See docs/INSTALL.md for the full matrix including uninstall and skill placement.
Quick start for an officially installed DSH:
# from npm (once published)
dsh plugin --profile web add dsh-daoing-memory
# or straight from GitHub
dsh plugin --profile web add github:daoing/dsh-daoing-memory
# place the extraction skill where DSH loads skills from
node node_modules/dsh-daoing-memory/scripts/install-skill.mjs
Then restart DSH. The memory tools become available to your agent, the profile snapshot starts being injected, and a Memory button appears at the foot of the web left sidebar (beside Settings) — clicking it opens the Memory workbench.
Usage
Once installed, the agent gains the memory_* tools. Typical flow:
- During a session the agent appends raw notes with
memory_fact.
- At a natural pause it runs
memory_extract to distill facts + concerns (guided by the memory-extraction skill).
- In later sessions
memory_recall surfaces relevant experiences; a compact profile snapshot is already present in the system prompt.
- When an experience is confirmed useful the agent calls
memory_report; when it is wrong, memory_revise.
- You can inspect and curate everything in the Memory workbench.
See docs/INSTALL.md for usage details and docs/STATUS.md for what is implemented today.
Design
The architecture, the trust/earning model, the data schema, and the anti-pollution boundaries are documented in docs/DESIGN.md. Current implementation status and extension directions live in docs/STATUS.md.
Project layout
dsh-daoing-memory/
├── lib/ # prebuilt artifacts (host + browser bundle + typert)
├── src/ # TypeScript source (for reference & iteration)
├── skill/ # bundled memory-extraction skill (standalone .md)
├── cordis.patch.yml # profile patch that wires the plugin into DSH
├── scripts/ # prepare + install-skill helpers
└── docs/ # INSTALL · DESIGN · STATUS · BUILDING · FAQ · MIGRATION
Building & publishing
This repository ships its build output (lib/) so that installing it never requires the DSH monorepo toolchain. See docs/BUILDING.md for how the package is produced, published to npm, and listed on the DSH plugin marketplace.
License
MIT © daoing
