dsh-chapters
Long memory for local AI: a DeepSeek Harness
plugin that archives agent conversations as verbatim chapters, continues sessions from a Table of
Contents at zero inference-token cost, and turns that archive into a shared, searchable knowledge
pool. Published on npm as @treeseed/dsh-chapters;
source at treeseed-ai/dsh-chapters.
All documentation lives on the documentation site — this README is a doorway, not a duplicate.
Install (the whole setup)
dsh plugin --profile web add @treeseed/dsh-chapters # install into your harness profile
dsh plugin --profile web update @treeseed/dsh-chapters # keep current — restart `dsh web` after
--profile <name> is required on every dsh plugin command (web = the browser harness profile).
After restart you have the fork button, seven agent tools, twelve /chapters-* commands (through
0.2.17, including /chapters-remove) — and the optional chapters preset that routes /compact
through the zero-token engine.
Find the right page for your question
| You want to… | Go to |
|---|
| understand what this is / why (the ten benefits) | Home → “What it gives you” |
| install, upgrade, uninstall, verify provenance | §23 Install, upgrade, provenance |
| run your first compaction/continuation | §1 Quickstart → §2 Your first continuation |
| link a shared knowledge pool (git or TreeDX) | §3 Link a knowledge pool · §10 The pool |
| run any command — exact syntax, scope, outputs | §15 Commands |
| delete a bad chapter (the wrong-project prompt) | §15 Commands → /chapters-remove |
| know what the agent-facing tools do | §16 Agent tools |
| tune a threshold, budget, or trigger | §17 Configuration |
| understand why not summaries / the economics | §5 Why not summaries? |
| read the design rules the system won't break | §6 The four invariants |
| the theory behind retrieval-over-compression (RLM) | §8 Recursive context theory |
| propose/approve rules for agents | §13 Rules & governance |
| let a second model propose your next prompt | §28 Prompt recommendations |
| debug something (“seems broken, is it?”) | §4 FAQ & troubleshooting |
| prompt-cache / llama.cpp behavior | §21 Cache behavior on llama.cpp |
| see measured results (production run + benchmarks) | §19 Benchmarks · §20 Field study |
| compare with other memory approaches | §22 The alternatives, compared |
| check security/privacy boundaries | §27 Security & privacy |
| release history & known-fixed incidents | §26 Changelog |
| develop the plugin itself | §24 Developing → the engineering specs below |
Developing / contributing
Three ways to set up a development environment — pick whichever fits your workflow:
| Mode | Best for | Command |
|---|
| Centralized (your live profile) | Day-to-day development on your main machine | ./dev/install/install.sh && dsh web |
| Compartmentalized (isolated DSH_HOME) | Testing changes without touching your live sessions | DSH_HOME=/tmp/dsh-dev ./dev/install/install.sh |
| Containerized (full stack) | Isolated platform with vLLM + Playwright + Chromium | cd dev/install && docker compose up -d |
All three install the same plugin set: dsh-chapters, dsh-codegraph
(structural code search via pre-indexed AST graph), and dsh-tinyfish-search. Full details in
dev/install/README.md. The containerized stack additionally ships its own
vLLM model server (Qwen3-4B-AWQ with inference + systemone endpoints), a DSH Web UI service, and a
CI-capable test runner — everything runs locally with no external API keys required.
The site's Part VI covers the daily loop; the repository itself holds the engineering truth for agents:
AGENTS.md — working spec for AI agents developing this repo (invariants, budget, hard rules)docs/ — deep specs: contract.md (host API facts), architecture.md (mechanics), development.md (build/test loop + tape system), verify.md (every claim's procedure), knowledge-repo.md (design record), host-compaction-seam.md, provider.mddev/install/ — installer, containerized platform, and local-llm skill- benchmarks:
npm run bench — three arms, local model, never CI (protocol)
Feature requests and bugs: open an issue —
measured reports get measured answers.
License
Apache-2.0 — see LICENSE. Your conversations, chapters, and knowledge pools stay yours.
