@achasoft/dsh-memory

Configuration

Every field is a validated setting, changeable from your profile's cordis.patch.yml or from Settings → Plugins → Memory in the Web Client.

Setting	Default	What it does
databasePath	.dsh/memory.db	Relative to each project's directory.
injectRules	true	Put the rule block in every model request. Off keeps rules stored but unenforced.
injectSessionContext	true	Seed a starting session with the last summary, sprint goals, and recent decisions.
autoSession	true	Open and close a memory session alongside each agent session.
remind	once	When to remind the model to file a summary: never, once, every-turn.
vectorWeight	0.6	The semantic signal's share of a blended ranking. Ignored without embeddings.
minSimilarity	0.05	Similarity floor for a search that does not name one.
searchLimit	10	Hits returned when the caller does not say.
candidateLimit	1000	Rows either search probe considers.
embedBatch	64	Memories embedded per background pass.
toolset	core	core or full — see below.
retentionDays	per category	Days per category; 0 means never. Rules never expire regardless.

Semantic recall (optional)

Enable the embeddings row and point it at any endpoint speaking OpenAI's /v1/embeddings — a hosted API, a local inference server, or Ollama:

- id: memory-embeddings-openai
  disabled: false
  config:
    baseUrl: https://api.openai.com/v1
    model: text-embedding-3-small
    apiKeyEnv: OPENAI_API_KEY
    timeoutMs: 30000
    batchSize: 64

The key is addressed by reference, never stored: apiKeyEnv names an environment variable resolved through the harness credential seam at the start of every call.

Memories are embedded in the background and vectors are stored as DuckDB FLOAT[]. Changing model strands the old vectors — they are excluded from comparison rather than compared — and Rebuild vectors in the UI re-embeds everything.