Global Settings
These settings control how Atlas stores, scores, and retains memories across
your entire namespace.
Scoring weights
The hybrid scoring function S = α·V + β·R + γ·F + δ·A uses these defaults:
| Parameter | Default | Description |
|---|
alpha (V) | 0.20 | Semantic similarity weight |
beta (R) | 0.70 | Recency weight — dominant signal |
gamma (F) | 0.10 | Access frequency weight |
delta (A) | 0.50 | Graph relation alignment weight |
scoring_weights in your retrieve payload:
{
"query": "What changed last week?",
"user_id": "user-123",
"scoring_weights": {
"alpha": 0.15,
"beta": 0.85,
"gamma": 0.05,
"delta": 0.30
}
}
Increase beta (recency) for agents that track fast-changing state
(e.g. project management, support tickets). Increase alpha for
agents doing document retrieval where recency matters less.
Semantic floor filter
Facts with semantic similarity below this threshold are dropped before scoring.
Prevents noise from unrelated graph traversal.
Default: V_min = 0.22
Lower it to surface more results. Raise it to tighten precision.
Memory consolidation schedule
Consolidation runs Ebbinghaus decay, prunes weak memories, and compresses
semantically similar clusters via LLM.
POST /brain/consolidate
{
"user_id": "user-123",
"force": false
}
| Parameter | Default | Description |
|---|
force | false | Run even if consolidation ran recently |
decay_rate | 0.85 | Confidence multiplier per consolidation cycle |
prune_threshold | 0.15 | Delete facts below this confidence |
compress_cluster_size | 5 | Min cluster size for LLM compression |
Retrieval defaults
{
"k": 5,
"include_episodic": true,
"include_semantic": true,
"max_hops": 2
}
max_hops for complex relational queries (e.g. "Who depends on X?").
Set include_episodic: false for faster graph-only retrieval.
Rate limits by plan
| Plan | Requests/min | Ops/month |
|---|
| Free | 10 rpm | 1,000 |
| Starter | 60 rpm | 50,000 |
| Pro | 300 rpm | 500,000 |
| Scale | 1,000 rpm | 5,000,000 |
HTTP 429 with "detail": "ops_limit_exceeded".
