The contextual-learning layer for AI agents. We ingest your production telemetry, score every run, and write what works into a memory your agents read before the next one.
A timeline of every reasoning call, tool invocation, and result — with full inputs, outputs, and timing. Replayable, queryable, framework-agnostic.
3c2a574b-ca03-44a4-ba72…{
"goal": "Read the latest email and reply with 'chicken' in it",
"iteration": 1,
"context": {
"lastAction": null,
"authenticated": false
}
}{
"action": "authenticate",
"reasoning": "The agent is not authenticated, so it needs to authenticate first."
}Configurable rubrics evaluate the run end-to-end. Spolm pinpoints what went right, what went wrong, and writes a root-cause summary traced to the exact step that broke.
Successful and failed decision patterns get written to a knowledge base ranked by accuracy. Before each run, the agent retrieves the most relevant lessons — and adapts strategy pre-emptively.
authenticate before fetch_emailDrop the SDK into your agent loop. Spolm tails your runs, scores them, and exposes a single retrieve() call that returns relevant context for the next prompt.
# 1. wrap your run from spolm import Spolm sp = Spolm(api_key="sk-…") # 2. retrieve learned context before each run ctx = sp.retrieve(task="reply to email") # 3. instrument the run with sp.run(task) as r: result = agent.invoke(task, context=ctx) r.log_result(result) # 4. spolm scores it, stores the lessons, ranks them # next retrieve() gets smarter automatically.
Catch reasoning loops, hand-off mistakes, and tone drift before they reach customers. Tighter every week, no retraining.
When one agent navigates a knowledge base for a query, the next inherits the path. Lower tokens, faster answers.
Naming, internal APIs, test idioms — learned once, applied across every PR. The agent stops re-introducing the same bugs.