dsh-experience-recall

中文 ｜ DeepSeek Harness plugin (profile bundle)

Your agent forgets what it learned in other projects. This plugin hands it back at the right moment.

It watches what the model is doing — its reasoning, its tool calls, its failures — and when the running work matches something your memory store already learned somewhere else, it injects that one lesson as a short message right before the next step. Everything runs locally: the keyword table, the retrieval and the relevance judge use your own machine by default. Injected messages still consume the host model's context and input tokens.

What it looks like

A real injection from a DSH session (trimmed):

【可能相关的历史经验 · 来自其他项目】
· 验证自研 DSH 宿主插件的一次性宿主法（不必重启 GUI、不碰 ~/.dsh）：把 DSH_HOME 指到 %TEMP% 里的临时 home，
  拷入 settings.yaml 与 .credentials.yaml，profile 的 dsh.profile.bundles 只写 [dsh-base, dsh-headless, 自研插件]，
  然后用 dsh --profile …（记忆 id: 0f5eddca）

需要细节：用 mnemon_recall / document_search 取全文。

The card is deliberately one line with the reusable conclusion, plus an id so the model can pull the full memory if it needs it. At most a few of these per turn, ~1.2 KB each.

How it works

session/event (reasoning · text · tool arguments · tool results · errors)
  → mechanical keyword match against a vocabulary derived from your memories   (free, no model)
  → gates: once per term per turn · once per memory · cooldown · per-turn quota
  → local small model judges "does this lesson change what I should do next?"  (serialized, ~0.4 s)
  → the ready result is appended at the next agent/pre-step boundary

The boundary never waits: if retrieval or judging is not finished when the next step starts, that step simply proceeds and the card arrives a step later. Nothing is injected on a guess — an unparsable or uncertain judgement is dropped rather than shown.

The vocabulary is not hand-written. It is derived from your memory store (author tags/entities plus the proper nouns in the text) and from your project documents, including each document's ## sections, then filtered for terms that cannot discriminate anything. New memories are picked up automatically after a mnemon_* tool call.

Requirements

Need	Why	Without it
DSH >= 0.1.5-rc.1	host	—
Node >= 22.5, 24.x recommended	node:sqlite reads the memory store. On Node 22 that module needs --experimental-sqlite; it is unflagged from Node 23.4	the plugin still runs, but the memory-derived vocabulary is empty (documents and seed terms keep working)
dsh-mnemon + the mnemon CLI	memory indexing and semantic search	without the CLI, already indexed cards can still trigger; CLI search is unavailable
A local model on Ollama (http://localhost:11434, default qwen3.5:9b)	builds the cards and judges relevance	it observes and logs, but injects nothing — it refuses to guess

Host compatibility. An injected message carries a DSH v4 producer-owned source (kind: "plugin:dsh-experience-recall", the same kind the host's own v3→v4 migration derives for a third-party plugin). Reading also recognises the pre-v4 { kind: "plugin", plugin } wrapper, so injections recorded by a 0.6.x / 0.7.0 session are still treated as plugin messages — for visibility, de-duplication, the re-injection distance and the window scan — and never as user input.

Install

dsh plugin --profile web add dsh-experience-recall

web is the profile behind the DSH web UI — substitute your own profile name if you renamed it. The command forwards to your package manager, so it also accepts a repository or a local checkout:

# track this repository instead of the npm release
dsh plugin --profile web add github:OverDustD7/dsh-experience-recall

# development: install the checkout you are working in
dsh plugin --profile web add link:/path/to/dsh-experience-recall

Then restart dsh web. Installing already adds the package to the profile's dsh.profile.bundles — do not add it by hand as well, or the loader fails with duplicate loader entry id.

Verify it works

• In chat: /experience-recall prints the vocabulary size, scan/hit counters, how many candidates the judge accepted, rejected or never answered (with average latency and retry count), injections, and the last card that got through.
• On disk: $DSH_HOME/dsh-experience-recall/logs/observe.ndjson records trigger / recall (with the judge's reason) / verify / inject / term-cooldown — "why did it inject / why didn't it" is always answerable. Verdict records also carry which half of the vocabulary fired (source/termKind) and how common that term is.
• Install self-check: node node_modules/dsh-experience-recall/tools/check-install.mjs (add --cache / --log if you moved the state directory).

Give it a few minutes of real work first: on a fresh install the first start builds one card per memory with the local model (a few hundred memories ≈ 5 minutes), and the vocabulary is thin until that finishes.

Switching the local small model

/experience-recall model (no argument) asks the local Ollama what chat models are installed and shows a picker card — the one in use comes first, the rest sorted smallest-first; embedding models such as bge-m3 are never offered. The choice applies immediately: the judge and the card builder both read the value per call, so the next memory uses the new model with no restart, and the value is also written back to your profile patch so the next restart keeps it.

/experience-recall model                    # list the local models and pick one
/experience-recall model qwen3.6:35b-a3b    # switch straight to it, no card
/experience-recall help                     # subcommand help

Use the named form when you would rather skip the card or Ollama is not running. A wrong name is safe: if the model is not installed the command says so ("not installed — the judge will fail and cards fall back to the mechanical builder"). Cards already built keep the previous model's wording — switching only affects cards built or rebuilt afterwards.

Configuration

Everything is optional and lives in the row's config — either in the package's cordis.patch.yml or, better, in your own profile patch ($DSH_HOME/profiles/<name>/cordis.patch.yml), which wins:

- id: experience-recall
  config:
    maxInjectionsPerTurn: 0     # keep observing, inject nothing
    localModel: 'qwen3.5:9b'

Key	Default	Meaning
enabled	true	master switch
stateDir	$DSH_HOME/dsh-experience-recall	where the log and the card cache live
logPath / cardCachePath	under stateDir	override individually; cardCachePath: '' disables the cache
mnemonCliPath	$MNEMON_CLI_PATH	the mnemon CLI; '' disables CLI retrieval only
mnemonDataDir / mnemonStore	$MNEMON_DATA_DIR or $DSH_HOME/mnemon / $MNEMON_STORE or default	shared by the index reader and the CLI; unset DSH_HOME uses ~/.dsh
localModel / localEndpoint	qwen3.5:9b / http://localhost:11434	card builder + relevance judge (this is what /experience-recall model changes)
modelListTimeoutMs / modelPickTimeoutMs	5000 / 300000	timeout for fetching the installed-model list; how long the picker card waits for the human
cardBuilder	local	mechanical skips the model when building cards
minScore	0.35	retrieval score floor
maxInjectionsPerTurn / maxCardsPerInjection / maxInjectBytes	3 / 2 / 1200	injection budget
verifyEnabled / verifyTimeoutMs / verifyMinIntervalMs	true / 6000 / 1200	the relevance gate
verifyRetryTimeoutMs / verifyRetryLimit	30000 / 1	one retry with a longer budget when the local model never answered — aborting a cold load cancels it in Ollama, so retrying inside the same budget livelocks that load
verifyBackoffBaseMs / verifyBackoffMaxMs	2000 / 30000	after an unanswered judgement the next start is spaced out (base * 2^n, capped) instead of burning one call per candidate
warmupAttempts / warmupRetryDelayMs	3 / 5000	how long startup keeps trying to load the local model before real judgements begin
documentFragments / fragmentChars	true / 200	index each document's ## sections as their own cards
systemPromptNote	true	register the constant explanation section in the system prompt
reinjectTokenDistance / repeatToolLimit / repeatWindowSteps	20000 / 3 / 12	long-session criteria: forgotten injections, going in circles
termRejectLimit / termBlockCooldownMs	3 / 1800000	term reputation: a term rejected this often without ever passing is cooled down

What it writes, and what it costs

• State only, under stateDir: logs/observe.ndjson (rotated at 32 MB) and cache/cards.json. Nothing is written inside node_modules, and no file of yours is modified.
• Retrieval and judging use the configured endpoints (localhost by default), with no telemetry. A custom endpoint receives the context/card or embedding query needed for its request. Logs contain keyword context and injected card text. The relevance judge runs as a local node subprocess.
• Context cost is bounded: at most maxInjectionsPerTurn injection blocks per turn, each ≤ maxInjectBytes, with at most maxCardsPerInjection cards per block. Injected text counts toward host model input tokens.

Uninstall

dsh plugin --profile web remove dsh-experience-recall

Then restart. Delete $DSH_HOME/dsh-experience-recall/ if you also want the log and cache gone.

Troubleshooting

Symptom	Likely cause
Injections never happen	check the store paths, injection quota, model and endpoint; an unavailable judge drops candidates instead of guessing. Built-in seed terms do not require the model
The log shows many rejected verdicts and few inject	the judge rejects weak or uncertain matches. /experience-recall and term-cooldown show term reputation; unanswered judgements do not count against a term
Nothing is injected for the first minutes after a restart	the local model is loading. Since 0.6.2 a timed-out judgement retries with a longer budget instead of cancelling that load
First start is slow	the card cache is being built: one local model call per memory. Prebuild it with node node_modules/dsh-experience-recall/tools/build-cards.mjs --cache <stateDir>/cache/cards.json
Nothing at all after install	the plugin needs a restart; check ready in the log and that command-registered / system-prompt-section / token-meter appear

Development

Layout:

Path	What is in it
lib/index.js	wiring: per-agent subscriptions to session/event and agent/pre-step, the /experience-recall command, the constant system-prompt section, the token meter
lib/models.js · lib/model-command.js	/experience-recall model: list the local models, rank them into options, switch and write the profile back (no host API, so it tests offline)
lib/observer.js · lib/scan.js	turn session events into scannable text segments
lib/keywords.js · lib/table.js · lib/terms.js	the vocabulary: matching, the live term table, term extraction and shape gates
lib/cards.js	builds cards (local model + mechanical fallback + document fragments) and owns the disk cache
lib/mnemon-store.js · lib/memory-watch.js	read the memory store (read-only node:sqlite) and reconcile it incrementally
lib/recall.js	the mnemon CLI search (hard timeout, cache, silent degradation)
lib/verify.js · tools/check-relevance.mjs	the second-stage relevance judge (serialized, retried, backed off)
lib/controller.js · lib/render.js · lib/surface.js	gates, queue, injection boundary; rendering; visibility checks
tools/	CLI tools: check-install, build-cards, term-audit, verdict-report, judge-probe, verify-package, and the local evaluation harness
test/	145 tests on node's built-in runner

Commands (no install step is needed — the plugin has no dependencies):

npm run check                            # syntax check every runtime module
npm test                                 # 145 unit tests, no network
node tools/check-install.mjs             # installation self-check
node tools/verify-package.mjs            # simulate the published package from `files`
node tools/verdict-report.mjs            # runtime data: sources, scores, judge health
node tools/judge-probe.mjs --n 3         # probe the judge against the real endpoint

lib/config.js holds PLUGIN_VERSION, which mirrors package.json; prepack runs the syntax check and the test suite, so a broken package cannot be published. CI runs the same two commands on Ubuntu and Windows with Node 24.

The labelled pairs used to calibrate the judge prompt are not in this repository: they are drawn from the author's private memory excerpts, so tools/judge-ab.mjs, tools/calibrate.mjs and tools/relevance-lab.mjs only exist in the author's checkout. Everything needed to run, test and package the plugin is here.

Links

• DeepSeek Harness — the host this plugin runs in
• dsh-mnemon — the memory store and mnemon CLI it reads
• awesome-dsh-plugin — the curated list behind the DSH marketplace; GitHub topic dsh-plugin

License

MIT