ragmemory/ragmemory.example.ini at main · kuchris/ragmemory · 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
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
[structured_memory]
# Copy this file to ragmemory.local.ini and fill in your real key.
# ragmemory.local.ini is ignored by git.
# api_key = YOUR_NVIDIA_API_KEY
model = minimaxai/minimax-m2.7
max_chars = 6000
max_tokens = 900

[llm]
# Keep structured extraction on NVIDIA by default. You can move compaction to
# another OpenAI-compatible provider first, because compact_text is lower risk.
structured_provider = nvidia
compact_provider = nvidia

[llm.nvidia]
# api_key = YOUR_NVIDIA_API_KEY
base_url = https://integrate.api.nvidia.com/v1
model = minimaxai/minimax-m2.7
api_style = openai_chat

[llm.opencode_go]
# api_key = YOUR_OPENCODE_GO_API_KEY
base_url = https://opencode.ai/zen/go/v1
# Start with an OpenAI-compatible /chat/completions model.
model = deepseek-v4-flash
api_style = openai_chat
# OpenCode Go deepseek-v4-flash otherwise spends tokens on reasoning_content.
thinking = disabled

[embedding]
# Keep the repo default lightweight. For a stronger local trial, set:
# provider = sentence_transformers
# model = BAAI/bge-small-en-v1.5
# device = cpu
# normalize_embeddings = true
provider = chroma_default

[compact]
# Phase 1 only: raw messages stay as source of truth; compact_text is cached beside them.
enable = True
model = minimaxai/minimax-m2.7
min_chars = 1500
max_chars = 30000
max_tokens = 1200
target_ratio = 0.35
mode = background

[obsidian.topics]
min_count = 2
allowlist = ragmemory, obsidian, codex-hooks, memory-decay
denylist = code_reference, config, decision, constraint, preference,
           open_question, chart, table, text, profile,
           python, powershell, javascript, typescript, bash,
           ini, yaml, json, markdown, sql, html, css

[topic_regroup]
# Optional worker job for LLM-curated Obsidian topics. The worker writes
# topic_taxonomy.json; export falls back to count-based topics when absent.
enable = true
max_tokens = 6000
# Only the top leaf-topic summaries are sent to the LLM. The remaining leaf
# topics still stay in Obsidian; they are just left ungrouped by this run.
max_input_topics = 150
# Reject regroup results with fewer groups when enough leaf topics exist.
min_groups = 10
# Override only the topic-regroup LLM call. Useful when global OpenCode Go
# thinking is enabled but taxonomy JSON should stay strict.
thinking = disabled

[obsidian.files]
# File/path hubs are opt-in because they can clutter the Obsidian graph.
enable = false

[recall]
# Hook-injected context size. Lower these if Codex spends too many tokens on memory.
context_token_budget = 500
retrieve_top_k = 3
structured_top_k = 2
recent_messages = 3
include_recent = true
include_structured = true

[mcp.tools]
# Recall is owned by the UserPromptSubmit hook when hooks are installed.
enable_recall = false
# Normal memory writes are owned by hooks by default. Keep MCP save disabled
# when hooks are installed to avoid duplicate writes.
enable_save = false
# Tombstone-only removal tools. Keep disabled until guardrails are verified.
enable_tombstone = true