graph-memory/cordis.patch.yml at main · adoresever/graph-memory · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
- insert:
    - id: graph-memory
      name: 'graph-memory/dsh'
      config:
        dbPath: !!js dshHomePath('graph-memory/graph-memory.db')
        extractionEnabled: true
        recallEnabled: true
        recallMaxNodes: 6
        # Optional cosine floor. Leave unset unless calibrated for the chosen
        # embedding model; ranked top-k is provider-neutral by default.
        semanticScoreThreshold: !!js "process.env.GRAPH_MEMORY_SEMANTIC_SCORE_THRESHOLD === undefined ? undefined : Number(process.env.GRAPH_MEMORY_SEMANTIC_SCORE_THRESHOLD)"
        # Automatic recall is the default path and needs no tool schema.
        # Set to search for gm_search, or all for administrative tools.
        assistantTools: none
        maintenanceInterval: 6
        # Extraction is a separate, lightweight workload. If these variables
        # are set, this route takes precedence over the foreground Agent model.
        llmProvider: !!js process.env.GRAPH_MEMORY_LLM_PROVIDER
        llmModel: !!js process.env.GRAPH_MEMORY_LLM_MODEL
        llmReasoningEffort: !!js process.env.GRAPH_MEMORY_LLM_REASONING_EFFORT
        llmMaxTokens: !!js "process.env.GRAPH_MEMORY_LLM_MAX_TOKENS ? Number(process.env.GRAPH_MEMORY_LLM_MAX_TOKENS) : undefined"
        # Each committed DSH turn is one extraction job containing only the
        # original user question and the final visible assistant answer.
        # Tool traffic and reasoning are never sent to the extraction model.
        # Durable events are separate from the compacted model surface. The
        # safe default never deletes raw messages. To opt in, first use
        # keep: referenced with dryRun: true and inspect gm_stats/gm_maintain.
        messageRetention:
          keep: all
          recentTurns: 0
          retentionDays: 0
          batchSize: 500
          dryRun: false
        # Graph Memory retains the newest completed user turns and replaces
        # the older model-surface prefix itself. This is model-free: no DSH
        # compaction/summarization request and no growing checkpoint.
        contextCompactionEnabled: true
        freshTurnCount: 5
        # Secrets stay outside the bundle and repository. Set these variables
        # in the environment that starts DSH to enable semantic vector recall.
        embedding:
          apiKeyEnv: !!js "process.env.GRAPH_MEMORY_EMBEDDING_API_KEY ? 'GRAPH_MEMORY_EMBEDDING_API_KEY' : undefined"
          baseURL: !!js process.env.GRAPH_MEMORY_EMBEDDING_BASE_URL
          model: !!js process.env.GRAPH_MEMORY_EMBEDDING_MODEL
          dimensions: !!js "process.env.GRAPH_MEMORY_EMBEDDING_DIMENSIONS ? Number(process.env.GRAPH_MEMORY_EMBEDDING_DIMENSIONS) : undefined"