memoplus4dsh

Configuration

Set under the plugin's config: in the profile's cordis.patch.yml:

Key	Default	Meaning
extraction	turn_end	turn_end extracts facts after every completed turn; off disables extraction
injection	true	Inject top-k relevant memories at the first step of each turn
injectTopK	8	Max memories injected per turn
injectMaxChars	2000	Character cap for the injected memory block
injectMaxQueryChars	4000	Skip retrieval+injection for longer user messages (document dumps, not queries)
tools	true	Register memory_search / memory_remember / memory_visualize / memory_status tools
progressBridge	true	Bridge goal/todo/schedule/plan progress events into the memory graph (M8)
stateDedup	true	Retrieval keeps only the newest bridge state event per entity+family; history stays in the graph
embedding	true	Local ONNX embeddings; failure degrades to keyword-only retrieval
embeddingModel	multilingual	multilingual = distiluse-base-multilingual-cased-v2 (512-dim, ~135MB first-download, 50+ languages incl. Chinese); english = all-MiniLM-L6-v2 (384-dim, ~23MB). Switching re-embeds stored vectors lazily
embeddingBackend	auto	auto = harrier sidecar (microsoft/harrier-oss-v1-0.6b, 1024-dim, multilingual, ~10ms/text CPU) when its python env has sentence-transformers, else ONNX encoder; onnx / harrier to force. Query-side uses the model's trained instruction prompt
embedPython	(nerPython or python3)	Python executable for the harrier embedding sidecar
hfBaseUrl	https://huggingface.co	Mirror base URL for the embedding model download
queryExpansion	true	LLM query expansion during retrieval + verbatim-quote query distillation for injection (1024-token/30s bounded calls, results cached on disk per query)
entityMergeLlm	true	LLM-adjudicated entity merge at extraction (embedding candidates + one bounded call per turn; only explicit sure merges)
supersedeLlm	true	LLM-adjudicated supersede detection (relation cardinality; older values marked supersededBy, history kept; re-mention guard + mark propagation)
nerAssist	true	NER candidate hints for extraction (detector chain: PyTorch sidecar → ONNX package → off)
nerPython	python3	Python executable for the NER sidecar (needs torch gliner stanza in that env; models auto-download on first use)

The plugin resolves python3 from the dsh process PATH — when dsh is launched from your shell it inherits that environment, so an interpreter that already has the packages works with zero configuration. If yours does not, scripts/setup-python.sh creates a dedicated venv (sentence-transformers + torch/gliner/stanza) and prints the exact nerPython / embedPython lines to paste into cordis.patch.yml. | dataDir | /memoplus4dsh | Plugin data directory (journal, snapshots, model cache, expansion cache) | | extractionProvider / extractionModel | session's own route | Override the model route used for extraction/expansion calls | | extractionMaxTokens | 8192 | Output cap for extraction calls (reasoning models need the headroom) | | extractionCallTimeoutMs | 120000 | Per-call timeout; a stalled endpoint fails fast into the retry queue | | extractionMaxRetries | 2 | Retries after the first attempt; the turn is then skipped and logged | | snapshotThreshold | 1000 | Journal ops between snapshot compactions |

Extraction consumes your configured model's API quota — set extraction: off to opt out.