Memory & Knowledge Base

Cross-session learning recall — automatic memorization and knowledge retrieval

Krawl maintains a persistent knowledge base across research sessions. Key findings from each research run are automatically extracted and stored, then recalled at the start of future research to avoid re-researching known facts.

How It Works

Automatic Memorization (Post-Research)

After every research session completes:

The report is sent to Haiku 4.5 (Tier 1, cheap/fast)
The model extracts 5–15 key facts/findings as structured learnings
Each learning gets content, topic tags, source query, and source URLs
Learnings are stored in PostgreSQL, scoped to the API key

Only specific, factual, reusable knowledge is extracted — not vague statements or opinions. Each learning must be self-contained (understandable without the original report context).

Automatic Recall (Pre-Research)

At the start of each research session:

The query is searched against the knowledge base (up to 15 results)
Matching memories are formatted into a "What you already know" context block
This context is injected into the research prompts

The recall prompt instructs the research engine to avoid re-researching known facts and focus on finding new information, updates, or details.

API Endpoints

Search Memories

curl -H "X-API-Key: your-key" \
  "https://api.krawl.sh/memory?q=ethereum%20merge&limit=10"

Response:

{
  "query": "ethereum merge",
  "count": 3,
  "memories": [
    {
      "id": "550e8400-...",
      "content": "Ethereum transitioned from Proof of Work to Proof of Stake on September 15, 2022, reducing energy consumption by ~99.95%.",
      "source_query": "What happened with the Ethereum merge?",
      "source_urls": ["https://ethereum.org/en/upgrades/merge/"],
      "tags": ["ethereum", "proof-of-stake", "merge"],
      "created_at": "2025-01-10T14:30:00Z",
      "result_id": "result-uuid-..."
    }
  ]
}

Knowledge Base Stats

curl -H "X-API-Key: your-key" \
  https://api.krawl.sh/memory/stats

Delete a Memory

curl -X DELETE -H "X-API-Key: your-key" \
  https://api.krawl.sh/memory/MEMORY_UUID

Response:

{"deleted": true, "id": "MEMORY_UUID"}

SSE Events

`memory_recall`

Emitted at the start of research when prior memories are found:

{
  "memories_count": 5,
  "memories": [
    {
      "content": "Ethereum merged to PoS in Sep 2022",
      "tags": ["ethereum", "proof-of-stake"]
    }
  ]
}

Memory Entry Structure

Each memory entry contains:

Field	Type	Description
`id`	string	UUID
`content`	string	The extracted fact/finding (1–2 sentences)
`source_query`	string	The original research query that produced this learning
`source_urls`	string[]	URLs of sources (up to 10)
`tags`	string[]	1–5 topic tags (lowercase)
`created_at`	string	ISO timestamp
`result_id`	string\|null	ID of the research result that produced this memory

Scoping

Memories are scoped to the API key. Each API key has its own isolated knowledge base. When MOGRA_API_KEY is not set (no auth), all memories share a global namespace.

Extraction Details

The memory extraction prompt instructs the model to:

Extract 5–15 key facts per report
Focus on data points, metrics, dates, names, relationships, and conclusions
Make each learning self-contained
Skip vague or opinion-based statements
Add 1–3 topic tags per learning

The extraction uses the first 8,000 characters of the report to stay within Haiku's optimal context window.

Memory & Knowledge Base

On this page