Session Persistence¶

CRP supports in-process knowledge accumulation via the Contextual Knowledge Fabric (CKF). Facts extracted during a crp.SDKClient() session are available to every subsequent ask()/complete() call in the same process.

Cross-process persistence is not yet implemented

The current SDKClient keeps CKF state in memory for the lifetime of the client. Saving a session to disk and reloading it in another process is on the roadmap. Today, persist knowledge by keeping the client open or by re-ingesting source material at startup.

Memory tiers¶

Tier	Name	Lifetime	Storage
1	Hot (Envelope)	Current window	In-memory
2	Warm (Session)	Current `SDKClient`	In-memory
3	Cold (CKF)	Persistent (planned)	SQLite + vectors + graph

Facts flow downward: Hot → Warm → Cold on session close (planned). Facts flow upward: Cold → Warm → Hot on relevance match (planned).

Example: multi-turn analysis¶

import crp

client = crp.SDKClient(provider="ollama", model="qwen3-4b")

# Reconnaissance turns
client.complete("Enumerate subdomains for acme.com")
client.complete("Scan discovered hosts for open ports")
client.complete("Identify running services and versions")

# Later turn uses accumulated facts
answer = client.ask(
    "Based on the discovered services, identify potential vulnerabilities",
    depth="thorough",
)
print(answer.text)
print(answer.sources)

Inspect session state¶

s = client.session()
print(s.id)
print(s.status())
print(s.fact_count, s.window_count)
print(client.storage.overview())

What is persisted today¶

Fact graph and warm state for the current SDKClient instance.
Audit trail via client.audit.export() and client.audit.verify().

What is planned¶

Named sessions loaded by session_id or app_id.
Cold storage flush/load across process restarts.
SQLite/Redis/S3 backend selection via crp.config.yaml.

Storage SDK Reference