microscope-memory 1.2.0

Pure binary cognitive memory engine with REST API bridge for LLM integration. Zero-JSON core, mmap-based, hierarchical, sub-microsecond search.
Documentation 
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
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145

# Benchmark index for the 60 resonance facts, with the semantic path ON.
#
# bench_config.toml is not what the published comparison runs. That one sets
# semantic_weight = 0.0 and provider = "mock", so the index it builds
# contains no real vectors and the vector path returns nothing -- a measurement
# taken on it reports a lexical-only system. This file is the same benchmark
# corpus with a real MiniLM index, which is what the FAISS and sqlite rows in
# docs/measurements/real_embedding_comparison.json are compared against.
#
# Build:  MICROSCOPE_CONFIG=bench_config_semantic.toml microscope-mem build
# Measure: python scripts/compare_baselines.py --config bench_config_semantic.toml
# Microscope Memory Configuration Example
# Copy to config.toml and adjust paths.

[paths]
# Directory containing memory layer JSON files
layers_dir = "./bench_layers"

# Output directory for binary index files
output_dir = "./bench_output_semantic"

# Temporary directory for processing
temp_dir = "./bench_tmp"

[index]
# Block size in characters (fixed viewport)
block_size = 256

# Maximum depth levels (0-8)
max_depth = 8

# Header size in bytes
header_size = 32

# Automatically consolidate the hot append log into the main index.
auto_rebuild = true

# Trigger consolidation after this many pending entries.
auto_rebuild_entries = 250

# Maximum source entries per layer. Zero preserves the complete history.
layer_retention_entries = 2000  # bound each layer file; oldest entries are trimmed past this
max_blocks = 0                   # 0 = unbounded; >0 evicts lowest-score blocks when the index exceeds this
protect_min_importance = 8       # never evict blocks with importance >= this
promote_energy_threshold = 0.35  # dream auto-promotes blocks whose recall energy reaches this

[search]
# Default number of results
default_k = 10

# Zoom weight for 4D search
zoom_weight = 2.0

# Keyword boost factor
keyword_boost = 0.1

# Semantic blending weight (0.0 = pure hash coords, 1.0 = pure embedding coords)
# NOTE: 0.0 disables semantic ranking in 'recall'. Set to > 0.0 to enable hybrid search.
semantic_weight = 1.0
semantic_rank_gain = 2.0
# Emotional bias warp strength (0.0 = disabled, 0.0-1.0 = warp toward emotional centroid)
emotional_bias_weight = 0.0

[memory_layers]
# Configure which cognitive layers to include
layers = [
    "long_term",
    "short_term",
    "associative",
    "emotional",
    "relational",
    "reflections",
    "crypto_chain",
    "echo_cache",
    "rust_state"
]

[performance]
# Use memory mapping
use_mmap = true

# Cache size in MB
cache_size = 64

# Number of parallel workers for building
build_workers = 4

# Use GPU acceleration for soft search (requires --features gpu)
use_gpu = false

# Zstd compression for data.bin (requires --features compression)
compression = false

# Query cache TTL in seconds
cache_ttl_secs = 300

[embedding]
# Provider: "mock" (hash-based), "candle" (BERT), or "onnx" (ONNX Runtime)
provider = "candle"

# Model name (for candle/onnx providers)
model = "sentence-transformers/all-MiniLM-L6-v2"

# Embedding dimension
dim = 384

# Max depth to embed (deeper = more vectors, more disk).
# NOTE: D5-D8 (words/tokens) are usually too small for meaningful embeddings.
max_depth = 4

[server]
# HTTP server port
port = 6060
# Shared master secret for inbound HTTP auth. Empty = no auth (single-user
# localhost default). Required when binding on a non-loopback host.
# Two modes use the same secret:
#   X-API-Key: <secret>                     — shared-key mode (trusted team)
#   Authorization: Bearer <user_id>.<hmac>  — per-user identity; generate with
#     `microscope-mem token <user_id>` — the verified user_id is bound to the
#     caller, so a user can never read another user's scoped memories.
api_key = ""

# CORS origin
cors_origin = "*"

[logging]
# Log level: trace, debug, info, warn, error
level = "info"

# Log file path (optional)
file = "microscope.log"

# ─── Federation (optional) ───────────────────────────
# Uncomment and configure to enable multi-index federated search.
# Each federated index gets its own section.
#
# [[federation.indices]]
# name = "project_a"
# config_path = "/path/to/project_a/config.toml"
# weight = 1.0
#
# [[federation.indices]]
# name = "project_b"
# config_path = "/path/to/project_b/config.toml"
# weight = 0.8