Skip to main content

​
Prerequisites

​
1. Encode a fact

Write a fact to the USER.gpu channel: 
curl -X POST https://api.spectralmemory.com/encode \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "user_id": "alice",
    "label": "USER.gpu",
    "value": "RTX5090"
  }'

{
  "status": "ok",
  "label": "USER.gpu",
  "value": "RTX5090",
  "channel": 3,
  "channel_count": 1
}
The fact is now encoded onto carrier channel 3 and persisted as a .mem file server-side.

​
2. Fetch the memory block

Retrieve the 512-token [MEM] block for system prompt injection: 
curl "https://api.spectralmemory.com/context?user_id=alice" \
  -H "Authorization: Bearer YOUR_API_KEY"

{
  "user_id": "alice",
  "fdm_block": "[MEM]magistrate yi GOP unprotected...[/MEM]",
  "channel_count": 1,
  "channel_labels": ["USER.gpu"],
  "plain_index": "USER.gpu=RTX5090"
}
Inject fdm_block into your agent’s system prompt. The Hermes3 reader model can decode it. Other models see an opaque token sequence.

​
3. Decode a fact

Read a specific channel value: 
curl "https://api.spectralmemory.com/decode?user_id=alice&label=USER.gpu" \
  -H "Authorization: Bearer YOUR_API_KEY"

{
  "user_id": "alice",
  "label": "USER.gpu",
  "value": "RTX5090",
  "source": "state",
  "latency_ms": 0
}
The default fast path reads from persistent state (~0ms). Pass use_model=true to verify via model inference instead.

​
4. Flush before compression

If your agent is about to compress its context, call /flush first: 
curl -X POST https://api.spectralmemory.com/flush \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"user_id": "alice"}'

{
  "status": "ok",
  "channel_count": 1,
  "lossless": true
}
A fresh [MEM] block is re-injected before compression fires. Facts survive into the next session intact.

​
Next steps

Install the Hermes plugin

Automate memory injection for every session.

Available channels

See all 40 labeled fact channels.